Alina Oprea

Harsh Chaudhari, John Abascal, Alina Oprea et al. · eth-zurich

Property inference attacks allow an adversary to extract global properties of the training dataset from a machine learning model. Such attacks have privacy implications for data owners sharing their datasets to train machine learning models. Several existing approaches for property inference attacks against deep neural networks have been proposed, but they all rely on the attacker training a large number of shadow models, which induces a large computational overhead. In this paper, we consider the setting of property inference attacks in which the attacker can poison a subset of the training dataset and query the trained target model. Motivated by our theoretical analysis of model confidences under poisoning, we design an efficient property inference attack, SNAP, which obtains higher attack success and requires lower amounts of poisoning than the state-of-the-art poisoning-based property inference attack by Mahloujifar et al. For example, on the Census dataset, SNAP achieves 34% higher success rate than Mahloujifar et al. while being 56.5x faster. We also extend our attack to infer whether a certain property was present at all during training and estimate the exact proportion of a property of interest efficiently. We evaluate our attack on several properties of varying proportions from four datasets and demonstrate SNAP's generality and effectiveness. An open-source implementation of SNAP can be found at https://github.com/johnmath/snap-sp23.

Nikhil Kandpal, Krishna Pillutla, Alina Oprea et al. · deepmind

Fine-tuning is a common and effective method for tailoring large language models (LLMs) to specialized tasks and applications. In this paper, we study the privacy implications of fine-tuning LLMs on user data. To this end, we consider a realistic threat model, called user inference, wherein an attacker infers whether or not a user's data was used for fine-tuning. We design attacks for performing user inference that require only black-box access to the fine-tuned LLM and a few samples from a user which need not be from the fine-tuning dataset. We find that LLMs are susceptible to user inference across a variety of fine-tuning datasets, at times with near perfect attack success rates. Further, we theoretically and empirically investigate the properties that make users vulnerable to user inference, finding that outlier users, users with identifiable shared features between examples, and users that contribute a large fraction of the fine-tuning data are most susceptible to attack. Based on these findings, we identify several methods for mitigating user inference including training with example-level differential privacy, removing within-user duplicate examples, and reducing a user's contribution to the training data. While these techniques provide partial mitigation of user inference, we highlight the need to develop methods to fully protect fine-tuned LLMs against this privacy risk.

John Abascal, Stanley Wu, Alina Oprea et al.

Transfer learning has become an increasingly popular technique in machine learning as a way to leverage a pretrained model trained for one task to assist with building a finetuned model for a related task. This paradigm has been especially popular for $\textit{privacy}$ in machine learning, where the pretrained model is considered public, and only the data for finetuning is considered sensitive. However, there are reasons to believe that the data used for pretraining is still sensitive, making it essential to understand how much information the finetuned model leaks about the pretraining data. In this work we propose a new membership-inference threat model where the adversary only has access to the finetuned model and would like to infer the membership of the pretraining data. To realize this threat model, we implement a novel metaclassifier-based attack, $\textbf{TMI}$, that leverages the influence of memorized pretraining samples on predictions in the downstream task. We evaluate $\textbf{TMI}$ on both vision and natural language tasks across multiple transfer learning settings, including finetuning with differential privacy. Through our evaluation, we find that $\textbf{TMI}$ can successfully infer membership of pretraining examples using query access to the finetuned model. An open-source implementation of $\textbf{TMI}$ can be found on GitHub: https://github.com/johnmath/tmi-pets24.

Matthew Jagielski, Stanley Wu, Alina Oprea et al.

A large body of research has shown that machine learning models are vulnerable to membership inference (MI) attacks that violate the privacy of the participants in the training data. Most MI research focuses on the case of a single standalone model, while production machine-learning platforms often update models over time, on data that often shifts in distribution, giving the attacker more information. This paper proposes new attacks that take advantage of one or more model updates to improve MI. A key part of our approach is to leverage rich information from standalone MI attacks mounted separately against the original and updated models, and to combine this information in specific ways to improve attack effectiveness. We propose a set of combination functions and tuning methods for each, and present both analytical and quantitative justification for various options. Our results on four public datasets show that our attacks are effective at using update information to give the adversary a significant advantage over attacks on standalone models, but also compared to a prior MI attack that takes advantage of model updates in a related machine-unlearning setting. We perform the first measurements of the impact of distribution shift on MI attacks with model updates, and show that a more drastic distribution shift results in significantly higher MI risk than a gradual shift. Our code is available at https://www.github.com/stanleykywu/model-updates.

Antonio Emanuele Cinà, Kathrin Grosse, Ambra Demontis et al.

The success of machine learning is fueled by the increasing availability of computing power and large training datasets. The training data is used to learn new models or update existing ones, assuming that it is sufficiently representative of the data that will be encountered at test time. This assumption is challenged by the threat of poisoning, an attack that manipulates the training data to compromise the model's performance at test time. Although poisoning has been acknowledged as a relevant threat in industry applications, and a variety of different attacks and defenses have been proposed so far, a complete systematization and critical review of the field is still missing. In this survey, we provide a comprehensive systematization of poisoning attacks and defenses in machine learning, reviewing more than 100 papers published in the field in the last 15 years. We start by categorizing the current threat models and attacks, and then organize existing defenses accordingly. While we focus mostly on computer-vision applications, we argue that our systematization also encompasses state-of-the-art attacks and defenses for other data modalities. Finally, we discuss existing resources for research in poisoning, and shed light on the current limitations and open research questions in this research field.

Galen Andrew, Peter Kairouz, Sewoong Oh et al.

Privacy estimation techniques for differentially private (DP) algorithms are useful for comparing against analytical bounds, or to empirically measure privacy loss in settings where known analytical bounds are not tight. However, existing privacy auditing techniques usually make strong assumptions on the adversary (e.g., knowledge of intermediate model iterates or the training data distribution), are tailored to specific tasks, model architectures, or DP algorithm, and/or require retraining the model many times (typically on the order of thousands). These shortcomings make deploying such techniques at scale difficult in practice, especially in federated settings where model training can take days or weeks. In this work, we present a novel "one-shot" approach that can systematically address these challenges, allowing efficient auditing or estimation of the privacy loss of a model during the same, single training run used to fit model parameters, and without requiring any a priori knowledge about the model architecture, task, or DP training algorithm. We show that our method provides provably correct estimates for the privacy loss under the Gaussian mechanism, and we demonstrate its performance on well-established FL benchmark datasets under several adversarial threat models.

Giorgio Severi, Matthew Jagielski, Gökberk Yar et al.

Federated learning is a popular strategy for training models on distributed, sensitive data, while preserving data privacy. Prior work identified a range of security threats on federated learning protocols that poison the data or the model. However, federated learning is a networked system where the communication between clients and server plays a critical role for the learning task performance. We highlight how communication introduces another vulnerability surface in federated learning and study the impact of network-level adversaries on training federated learning models. We show that attackers dropping the network traffic from carefully selected clients can significantly decrease model accuracy on a target population. Moreover, we show that a coordinated poisoning campaign from a few clients can amplify the dropping attacks. Finally, we develop a server-side defense which mitigates the impact of our attacks by identifying and up-sampling clients likely to positively contribute towards target accuracy. We comprehensively evaluate our attacks and defenses on three datasets, assuming encrypted communication channels and attackers with partial visibility of the network.

Georgios Syros, Gokberk Yar, Simona Boboila et al.

Most machine learning applications rely on centralized learning processes, opening up the risk of exposure of their training datasets. While federated learning (FL) mitigates to some extent these privacy risks, it relies on a trusted aggregation server for training a shared global model. Recently, new distributed learning architectures based on Peer-to-Peer Federated Learning (P2PFL) offer advantages in terms of both privacy and reliability. Still, their resilience to poisoning attacks during training has not been investigated. In this paper, we propose new backdoor attacks for P2PFL that leverage structural graph properties to select the malicious nodes, and achieve high attack success, while remaining stealthy. We evaluate our attacks under various realistic conditions, including multiple graph topologies, limited adversarial visibility of the network, and clients with non-IID data. Finally, we show the limitations of existing defenses adapted from FL and design a new defense that successfully mitigates the backdoor attacks, without an impact on model accuracy.

Talha Ongun, Simona Boboila, Alina Oprea et al.

The cyber-threat landscape has evolved tremendously in recent years, with new threat variants emerging daily, and large-scale coordinated campaigns becoming more prevalent. In this study, we propose CELEST (CollaborativE LEarning for Scalable Threat detection, a federated machine learning framework for global threat detection over HTTP, which is one of the most commonly used protocols for malware dissemination and communication. CELEST leverages federated learning in order to collaboratively train a global model across multiple clients who keep their data locally, thus providing increased privacy and confidentiality assurances. Through a novel active learning component integrated with the federated learning technique, our system continuously discovers and learns the behavior of new, evolving, and globally-coordinated cyber threats. We show that CELEST is able to expose attacks that are largely invisible to individual organizations. For instance, in one challenging attack scenario with data exfiltration malware, the global model achieves a three-fold increase in Precision-Recall AUC compared to the local model. We also design a poisoning detection and mitigation method, DTrust, specifically designed for federated learning in the collaborative threat detection domain. DTrust successfully detects poisoning clients using the feedback from participating clients to investigate and remove them from the training process. We deploy CELEST on two university networks and show that it is able to detect the malicious HTTP communication with high precision and low false positive rates. Furthermore, during its deployment, CELEST detected a set of previously unknown 42 malicious URLs and 20 malicious domains in one day, which were confirmed to be malicious by VirusTotal.

Lisa Oakley, Sam Stites, Cameron Moy et al.

The membership inference problem for publicly released statistics from a private dataset is well-studied. When developing and formally analyzing attack strategies, however, the focus has been on attacks that model the population using only its marginals. In practice, these attacks can perform well on various populations, however most formal analysis is for populations that follow a product distribution. These strategies may fail to leverage useful information about the population that is important for understanding a realistic privacy threat. In this work, we explore the impact of providing an attacker with additional information about the attribute dependency structure of the population, motivated by examples where multiple parties may have access to similarly structured data, for example the US Census and the IRS. To model this scenario, we re-frame the membership inference problem with respect to a population represented as a Bayesian network (BN). We develop a framework based on Bayesian decision-making which can incorporate prior information about the population to launch more effective, specialized attacks. To evaluate our framework, we introduce a specific attack instantiation which computes the Bayesian posterior using a probabilistic program, and prove its equivalence to an optimal variant of the likelihood ratio test attack for two populations with strong attribute dependency. We implement our program in the Roulette probabilistic programming language and show experimentally that it outperforms the likelihood ratio test and inner product attacks on five commonly used BNs, where the population dependency structure is too complex for the existing attacks to be manually adapted.

Harsh Chaudhari, Matthew Jagielski, Alina Oprea

Secure multiparty computation (MPC) has been proposed to allow multiple mutually distrustful data owners to jointly train machine learning (ML) models on their combined data. However, by design, MPC protocols faithfully compute the training functionality, which the adversarial ML community has shown to leak private information and can be tampered with in poisoning attacks. In this work, we argue that model ensembles, implemented in our framework called SafeNet, are a highly MPC-amenable way to avoid many adversarial ML attacks. The natural partitioning of data amongst owners in MPC training allows this approach to be highly scalable at training time, provide provable protection from poisoning attacks, and provably defense against a number of privacy attacks. We demonstrate SafeNet's efficiency, accuracy, and resilience to poisoning on several machine learning datasets and models trained in end-to-end and transfer learning scenarios. For instance, SafeNet reduces backdoor attack success significantly, while achieving $39\times$ faster training and $36 \times$ less communication than the four-party MPC framework of Dalskov et al. Our experiments show that ensembling retains these benefits even in many non-iid settings. The simplicity, cheap setup, and robustness properties of ensembling make it a strong first choice for training ML models privately in MPC.

Giorgio Severi, Simona Boboila, Alina Oprea et al.

As machine learning (ML) classifiers increasingly oversee the automated monitoring of network traffic, studying their resilience against adversarial attacks becomes critical. This paper focuses on poisoning attacks, specifically backdoor attacks, against network traffic flow classifiers. We investigate the challenging scenario of clean-label poisoning where the adversary's capabilities are constrained to tampering only with the training data - without the ability to arbitrarily modify the training labels or any other component of the training process. We describe a trigger crafting strategy that leverages model interpretability techniques to generate trigger patterns that are effective even at very low poisoning rates. Finally, we design novel strategies to generate stealthy triggers, including an approach based on generative Bayesian network models, with the goal of minimizing the conspicuousness of the trigger, and thus making detection of an ongoing poisoning campaign more challenging. Our findings provide significant insights into the feasibility of poisoning attacks on network traffic classifiers used in multiple scenarios, including detecting malicious communication and application classification.

Luze Sun, Anshuman Suri, Harsh Chaudhari et al.

When practitioners fine-tune LLMs on unvetted datasets, an adversary can exploit the data supply chain through task-level poisoning: inserting a small number of crafted instruction-response pairs that cause the model to embed attacker-specified entities, such as a country, in outputs for a targeted task family while behaving normally elsewhere. We introduce PoisonForge, a benchmark that parameterizes this threat along four dimensions (bias type, poisoning mode, appearance count, and target output length) and evaluates 12 open-weight models (from 2B to 32B parameters) across five families under a primarily 1% poison budget. With only 10 poisoned examples among 1,000 fine-tuning examples, 11 of 12 models exceed a 70% attack success rate (ASR) in their most vulnerable configuration. Meanwhile, unintended leakage to non-target tasks remains below 0.5%, and models perform well on standard benchmarks. We analyze in detail the factors contributing to attack success. We observe that multiple appearances of an entity increase the ASR, the optimal poisoning mode depends on the semantic structure of the target entity, and ASR drops monotonically with the task output length. A correlation analysis and risk prediction model confirm that poisoning design choices, rather than model scale, are the primary causes of attack success, and that these patterns generalize to predict attack success on new tasks. We release all configurations, pipelines, and analysis code to support reproducible comparisons.

Evan Rose, Hidde Lycklama, Harsh Chaudhari et al.

Privacy-preserving machine learning (PPML) systems enable multiple data owners to collaboratively train models without revealing their raw, sensitive data by leveraging cryptographic protocols such as secure multi-party computation (MPC). While PPML offers strong privacy guarantees, it also introduces new attack surfaces: malicious data owners can inject poisoned data into the training process without being detected, thus undermining the integrity of the learned model. Although recent defenses, such as private input validation within MPC, can mitigate some specific poisoning strategies, they remain insufficient, particularly in preventing stealthy or distributed attacks. As the robustness of PPML remains an open challenge, strengthening trust in these systems increasingly necessitates post-hoc auditing mechanisms that instill accountability. In this paper we present UTrace, a framework for user-level traceback in PPML that attributes integrity failures to responsible data owners without compromising the privacy guarantees of MPC. UTrace encapsulates two mechanisms: a gradient similarity method that identifies suspicious update patterns linked to poisoning, and a user-level unlearning technique that quantifies each user's marginal influence on model behavior. Together, these methods allow UTrace to attribute model misbehavior to specific users with high precision. We implement UTrace within an MPC-compatible training and auditing pipeline and evaluate its effectiveness on four datasets spanning vision, text, and malware. Across ten canonical poisoning attacks, UTrace consistently achieves high detection accuracy with low false positive rates.

Harsh Chaudhari, Giorgio Severi, Alina Oprea et al.

The integration of machine learning (ML) in numerous critical applications introduces a range of privacy concerns for individuals who provide their datasets for model training. One such privacy risk is Membership Inference (MI), in which an attacker seeks to determine whether a particular data sample was included in the training dataset of a model. Current state-of-the-art MI attacks capitalize on access to the model's predicted confidence scores to successfully perform membership inference, and employ data poisoning to further enhance their effectiveness. In this work, we focus on the less explored and more realistic label-only setting, where the model provides only the predicted label on a queried sample. We show that existing label-only MI attacks are ineffective at inferring membership in the low False Positive Rate (FPR) regime. To address this challenge, we propose a new attack Chameleon that leverages a novel adaptive data poisoning strategy and an efficient query selection method to achieve significantly more accurate membership inference than existing label-only attacks, especially at low FPRs.

Giorgio Severi, Simona Boboila, John Holodnak et al.

The training phase of machine learning models is a delicate step, especially in cybersecurity contexts. Recent research has surfaced a series of insidious training-time attacks that inject backdoors in models designed for security classification tasks without altering the training labels. With this work, we propose new techniques that leverage insights in cybersecurity threat models to effectively mitigate these clean-label poisoning attacks, while preserving the model utility. By performing density-based clustering on a carefully chosen feature subspace, and progressively isolating the suspicious clusters through a novel iterative scoring procedure, our defensive mechanism can mitigate the attacks without requiring many of the common assumptions in the existing backdoor defense literature. To show the generality of our proposed mitigation, we evaluate it on two clean-label model-agnostic attacks on two different classic cybersecurity data modalities: network flows classification and malware classification, using gradient boosting and neural network models.

Xavier Cadet, Aditya Vikram Singh, Harsh Mamania et al.

Investigating cybersecurity incidents requires collecting and analyzing evidence from multiple log sources, including intrusion detection alerts, network traffic records, and authentication events. This process is labor-intensive: analysts must sift through large volumes of data to identify relevant indicators and piece together what happened. We present a RAG-based system that performs security incident analysis through targeted query-based filtering and LLM semantic reasoning. The system uses a query library with associated MITRE ATT\&CK techniques to extract indicators from raw logs, then retrieves relevant context to answer forensic questions and reconstruct attack sequences. We evaluate the system with five LLM providers on malware traffic incidents and multi-stage Active Directory attacks. We find that LLM models have different performance and tradeoffs, with Claude Sonnet~4 and DeepSeek~V3 achieving 100\% recall across all four malware scenarios, while DeepSeek costs 15$\times$ less (\$0.008 vs.\ \$0.12 per analysis). Attack step detection on Active Directory scenarios reaches 100\% precision and 82\% recall. Ablation studies confirm that a RAG architecture is essential: LLM baselines without RAG-enhanced context correctly identify victim hosts but miss all attack infrastructure including malicious domains and command-and-control servers. These results demonstrate that combining targeted query-based filtering with RAG-based retrieval enables accurate, cost-effective security analysis within LLM context limits.

Ali Naseh, Yuefeng Peng, Anshuman Suri et al.

Text-to-image (T2I) models are increasingly popular, producing a large share of AI-generated images online. To compare model quality, voting-based leaderboards have become the standard, relying on anonymized model outputs for fairness. In this work, we show that such anonymity can be easily broken. We find that generations from each T2I model form distinctive clusters in the image embedding space, enabling accurate deanonymization without prompt control or training data. Using 22 models and 280 prompts (150K images), our centroid-based method achieves high accuracy and reveals systematic model-specific signatures. We further introduce a prompt-level distinguishability metric and conduct large-scale analyses showing how certain prompts can lead to near-perfect distinguishability. Our findings expose fundamental security flaws in T2I leaderboards and motivate stronger anonymization defenses.

Harsh Chaudhari, Ethan Rathbun, Hanna Foerster et al.

Chain-of-Thought (CoT) reasoning has emerged as a powerful technique for enhancing large language models' capabilities by generating intermediate reasoning steps for complex tasks. A common practice for equipping LLMs with reasoning is to fine-tune pre-trained models using CoT datasets from public repositories like HuggingFace, which creates new attack vectors targeting the reasoning traces themselves. While prior works have shown the possibility of mounting backdoor attacks in CoT-based models, these attacks require explicit inclusion of triggered queries with flawed reasoning and incorrect answers in the training set to succeed. Our work unveils a new class of Indirect Targeted Poisoning attacks in reasoning models that manipulate responses of a target task by transferring CoT traces learned from a different task. Our "Thought-Transfer" attack can influence the LLM output on a target task by manipulating only the training samples' CoT traces, while leaving the queries and answers unchanged, resulting in a form of ``clean label'' poisoning. Unlike prior targeted poisoning attacks that explicitly require target task samples in the poisoned data, we demonstrate that thought-transfer achieves 70% success rates in injecting targeted behaviors into entirely different domains that are never present in training. Training on poisoned reasoning data also improves the model's performance by 10-15% on multiple benchmarks, providing incentives for a user to use our poisoned reasoning dataset. Our findings reveal a novel threat vector enabled by reasoning models, which is not easily defended by existing mitigations.

Ruben Chocron, Doron Jonathan Ben Chayim, Eyal Lenga et al.

AI agents are increasingly deployed to act autonomously in the world, yet there is still no reliable way to trace a harmful agent back to the account that deployed it. This creates the same accountability gap across both ends of the intent spectrum: benign operators may deploy misconfigured or overbroad agents that cause harm unintentionally, while malicious operators may deliberately weaponize agents for scams, harassment, or cyber attacks. In many cases, these agents are powered by vendor-hosted models, a dependency that holds even for sophisticated adversaries such as state actors conducting cyber operations. In either case, affected parties can observe the behavior but cannot notify the responsible operator, stop the session, or identify the account for investigation. We formalize this gap as the problem of agent attribution: linking an observed agent interaction to the responsible account at the hosting vendor. To our knowledge, this is the first work to define the problem and present a practical solution. Our protocol is canary-based: an authorized party injects a canary into the agent's interaction stream, and the vendor searches a narrow window of session logs to recover the originating session and account. Simple canaries suffice in non-adversarial settings. For adversarial operators who filter or paraphrase incoming content, we develop robust canary constructions that cannot be suppressed without degrading the agent's own task performance, yielding a formal asymmetry in the defender's favor. We evaluate a variety of scenarios including real-world agents and show that our attribution method is reliable, robust, and scalable for vendor-side deployment.

Evan Rose, Tushin Mallick, Matthew D. Laws et al.

Autonomous multi-agent systems based on large language models (LLMs) have demonstrated remarkable abilities in independently solving complex tasks in a wide breadth of application domains. However, these systems hit critical reasoning, coordination, and computational scaling bottlenecks as the size and complexity of their tasks grow. These limitations hinder multi-agent systems from achieving high-throughput processing for highly parallelizable tasks, despite the availability of parallel computing and reasoning primitives in the underlying LLMs. We introduce the Agent-Parallel Workload Architecture (APWA), a distributed multi-agent system architecture designed for the efficient processing of heavily parallelizable agentic workloads. APWA facilitates parallel execution by decomposing workflows into non-interfering subproblems that can be processed using independent resources without cross-communication. It supports heterogeneous data and parallel processing patterns, and it accommodates tasks from a wide breadth of domains. In our evaluation, we demonstrate that APWA can dynamically decompose complex queries into parallelizable workflows and scales on larger tasks in settings where prior systems fail completely.

Matthew D. Laws, Alina Oprea, Cristina Nita-Rotaru

Agentic AI governance is a critical component of agentic AI infrastructure ensuring that agents follow their owner's communication and interaction policies, and providing protection against attacks from malicious agents. The state-of-the-art solution, SAGA, assumes a logically centralized point of trust, the Provider, which serves as a repository for user and agent information and actively enforces policies. While SAGA provides protection against malicious agents, it remains vulnerable to a malicious Provider that deviates from the protocol, undermining the security of the identity and access control infrastructure. Deployment on both private and public clouds, each susceptible to insider threats, further increases the risk of Provider compromise. In this work, we analyze the attacks that can be mounted from a compromised Provider, taking into account the different system components and realistic deployments. We identify and execute several concrete attacks with devastating effects: undermining agent attributability, extracting private data, or bypassing access control. We then present three types of solutions for securing the Provider that offer different trade-offs between security and performance. We first present SAGA-BFT, a fully byzantine-resilient architecture that provides the strongest protection, but incurs significant performance degradation, due to the high-cost of byzantine resilient protocols. We then propose SAGA-MON and SAGA-AUD, two novel solutions that leverage lightweight server-side monitoring or client-side auditing to provide protection against most classes of attacks with minimal overhead. Finally, we propose SAGA-HYB, a hybrid architecture that combines byzantine-resilience with monitoring and auditing to trade-off security for performance. We evaluate all the architectures and compare them with SAGA. We discuss which solution is best and under what conditions.

Sae Furukawa, Alina Oprea

Supervised Finetuning (SFT) has become one of the primary methods for adapting a large language model (LLM) with extensive pre-trained knowledge to domain-specific, instruction-following tasks. SFT datasets, composed of instruction-response pairs, often include user-provided information that may contain sensitive data such as personally identifiable information (PII), raising privacy concerns. This paper studies the problem of PII reconstruction from SFT models for the first time. We construct multi-turn, user-centric Q&A datasets in sensitive domains, specifically medical and legal settings, that incorporate PII to enable realistic evaluation of leakage. Using these datasets, we evaluate the extent to which an adversary, with varying levels of knowledge about the fine-tuning dataset, can infer sensitive information about individuals whose data was used during SFT. In the reconstruction setting, we propose COVA, a novel decoding algorithm to reconstruct PII under prefix-based attacks, consistently outperforming existing extraction methods. Our results show that even partial attacker knowledge can significantly improve reconstruction success, while leakage varies substantially across PII types.

Sepideh Avizeh, Tushin Mallick, Alina Oprea et al.

Our computing ecosystem is being transformed by two emerging paradigms: the increased deployment of agentic AI systems and advancements in quantum computing. With respect to agentic AI systems, one of the most critical problems is creating secure governing architectures that ensure agents follow their owners' communication and interaction policies and can be held accountable for the messages they exchange with other agents. With respect to quantum computing, existing systems must be retrofitted and new cryptographic mechanisms must be designed to ensure long-term security and quantum resistance. In fact, NIST recommends that standard public-key cryptographic algorithms, including RSA, Diffie-Hellman (DH), and elliptic-curve constructions (ECC), be deprecated starting in 2030 and disallowed after 2035. In this paper, we present MAGIQ, a framework for policy definition and enforcement in multi-agent AI systems using novel, highly efficient, quantum-resistant cryptographic protocols with proven security guarantees. MAGIQ (i) allows users to define rich communication and access-control policy budgets for agent-to-agent sessions and tasks, including global budgets for one-to-many agent sessions; (ii) enforces such policies using post-quantum cryptographic primitives; (iii) supports session-based enforcement of policies for agent-to-agent and one-to-many agent sessions; and (iv) provides accountability of agents to their users through message attribution. We formally model and prove the correctness and security of the system using the Universal Composability (UC) framework. We evaluate the computation and communication overhead of our framework and compare it with the state-of-the-art agentic AI framework SAGA. MAGIQ is a first step toward post-quantum-secure solutions for agentic AI systems.

Luze Sun, Alina Oprea, Eric Wong

LLM-based vulnerability detectors are increasingly deployed in security-critical code review, yet their resilience to evasion under behavior-preserving edits remains poorly understood. We evaluate detection-time integrity under a semantics-preserving threat model by instantiating diverse behavior-preserving code transformations on a unified C/C++ benchmark (N=5000), and introduce a metric of joint robustness across different attack methods/carriers. Across models, we observe a systemic failure of semantic invariant adversarial transformations: even state-of-the-art vulnerability detectors perform well on clean inputs while predictions flip under behavior-equivalent edits. Universal adversarial strings optimized on a single surrogate model remain effective when transferred to black-box APIs, and gradient access can further amplify evasion success. These results show that even high-performing detectors are vulnerable to low-cost, semantics-preserving evasion. Our carrier-based metrics provide practical diagnostics for evaluating LLM-based code detectors.

Shuyi Lin, Anshuman Suri, Alina Oprea et al.

LLM agents emit actions, not just text, and once taken, those actions often cannot be undone. Yet today's agent-safety evaluations run greedy or a few sampled rollouts and report a single safe/unsafe rate -- blind to the long-tail trajectories where unsafe behavior may arise from low-probability but non-negligible actions. We argue agent safety should be measured by search, not sampling. We apply BOA, a framework that, given a deployment configuration (model, decoder, prompt, environment, judger, likelihood budget), searches the in-budget trajectory space and reports a safety score: the probability the agent stays safe under the configuration. BOA searches both within a single LLM round and across the agent-environment interaction tree under a given likelihood budget, and makes search practical via batched decoding/judging, prefix caching, and chunked tree expansion. On agent-safety workloads, BOA discovers unsafe trajectories that greedy and sampled evaluations miss. BOA can additionally be used for ranking models, defenses, and attacks, all on the same scale, with manageable GPU costs.

Aditya Vikram Singh, Ethan Rathbun, Emma Graham et al.

Recent advances in multi-agent reinforcement learning (MARL) have created opportunities to solve complex real-world tasks. Cybersecurity is a notable application area, where defending networks against sophisticated adversaries remains a challenging task typically performed by teams of security operators. In this work, we explore novel MARL strategies for building autonomous cyber network defenses that address challenges such as large policy spaces, partial observability, and stealthy, deceptive adversarial strategies. To facilitate efficient and generalized learning, we propose a hierarchical Proximal Policy Optimization (PPO) architecture that decomposes the cyber defense task into specific sub-tasks like network investigation and host recovery. Our approach involves training sub-policies for each sub-task using PPO enhanced with cybersecurity domain expertise. These sub-policies are then leveraged by a master defense policy that coordinates their selection to solve complex network defense tasks. Furthermore, the sub-policies can be fine-tuned and transferred with minimal cost to defend against shifts in adversarial behavior or changes in network settings. We conduct extensive experiments using CybORG Cage 4, the state-of-the-art MARL environment for cyber defense. Comparisons with multiple baselines across different adversaries show that our hierarchical learning approach achieves top performance in terms of convergence speed, episodic return, and several interpretable metrics relevant to cybersecurity, including the fraction of clean machines on the network, precision, and false positives.

Ali Naseh, Yuefeng Peng, Anshuman Suri et al.

Retrieval-Augmented Generation (RAG) enables Large Language Models (LLMs) to generate grounded responses by leveraging external knowledge databases without altering model parameters. Although the absence of weight tuning prevents leakage via model parameters, it introduces the risk of inference adversaries exploiting retrieved documents in the model's context. Existing methods for membership inference and data extraction often rely on jailbreaking or carefully crafted unnatural queries, which can be easily detected or thwarted with query rewriting techniques common in RAG systems. In this work, we present Interrogation Attack (IA), a membership inference technique targeting documents in the RAG datastore. By crafting natural-text queries that are answerable only with the target document's presence, our approach demonstrates successful inference with just 30 queries while remaining stealthy; straightforward detectors identify adversarial prompts from existing methods up to ~76x more frequently than those generated by our attack. We observe a 2x improvement in TPR@1%FPR over prior inference attacks across diverse RAG configurations, all while costing less than $0.02 per document inference.

Evan Li, Tushin Mallick, Evan Rose et al.

LLM-integrated app systems extend the utility of Large Language Models (LLMs) with third-party apps that are invoked by a system LLM using interleaved planning and execution phases to answer user queries. These systems introduce new attack vectors where malicious apps can cause integrity violation of planning or execution, availability breakdown, or privacy compromise during execution. In this work, we identify new attacks impacting the integrity of planning, as well as the integrity and availability of execution in LLM-integrated apps, and demonstrate them against IsolateGPT, a recent solution designed to mitigate attacks from malicious apps. We propose Abstract-Concrete-Execute (ACE), a new secure architecture for LLM-integrated app systems that provides security guarantees for system planning and execution. Specifically, ACE decouples planning into two phases by first creating an abstract execution plan using only trusted information, and then mapping the abstract plan to a concrete plan using installed system apps. We verify that the plans generated by our system satisfy user-specified secure information flow constraints via static analysis on the structured plan output. During execution, ACE enforces data and capability barriers between apps, and ensures that the execution is conducted according to the trusted abstract plan. We show experimentally that ACE is secure against attacks from the InjecAgent and Agent Security Bench benchmarks for indirect prompt injection, and our newly introduced attacks. We also evaluate the utility of ACE in realistic environments, using the Tool Usage suite from the LangChain benchmark. Our architecture represents a significant advancement towards hardening LLM-based systems using system security principles.

Georgios Syros, Anshuman Suri, Jacob Ginesin et al.

Large Language Model (LLM)-based agents increasingly interact, collaborate, and delegate tasks to one another autonomously with minimal human interaction. Industry guidelines for agentic system governance emphasize the need for users to maintain comprehensive control over their agents, mitigating potential damage from malicious agents. Several proposed agentic system designs address agent identity, authorization, and delegation, but remain purely theoretical, without concrete implementation and evaluation. Most importantly, they do not provide user-controlled agent management. To address this gap, we propose SAGA, a scalable Security Architecture for Governing Agentic systems, that offers user oversight over their agents' lifecycle. In our design, users register their agents with a central entity, the Provider, that maintains agent contact information, user-defined access control policies, and helps agents enforce these policies on inter-agent communication. We introduce a cryptographic mechanism for deriving access control tokens, that offers fine-grained control over an agent's interaction with other agents, providing formal security guarantees. We evaluate SAGA on several agentic tasks, using agents in different geolocations, and multiple on-device and cloud LLMs, demonstrating minimal performance overhead with no impact on underlying task utility in a wide range of conditions. Our architecture enables secure and trustworthy deployment of autonomous agents, accelerating the responsible adoption of this technology in sensitive environments.

Harsh Chaudhari, Jamie Hayes, Matthew Jagielski et al. · deepmind

Model distillation has become essential for creating smaller, deployable language models that retain larger system capabilities. However, widespread deployment raises concerns about resilience to adversarial manipulation. This paper investigates vulnerability of distilled models to adversarial injection of biased content during training. We demonstrate that adversaries can inject subtle biases into teacher models through minimal data poisoning, which propagates to student models and becomes significantly amplified. We propose two propagation modes: Untargeted Propagation, where bias affects multiple tasks, and Targeted Propagation, focusing on specific tasks while maintaining normal behavior elsewhere. With only 25 poisoned samples (0.25% poisoning rate), student models generate biased responses 76.9% of the time in targeted scenarios - higher than 69.4% in teacher models. For untargeted propagation, adversarial bias appears 6x-29x more frequently in student models on unseen tasks. We validate findings across six bias types (targeted advertisements, phishing links, narrative manipulations, insecure coding practices), various distillation methods, and different modalities spanning text and code generation. Our evaluation reveals shortcomings in current defenses - perplexity filtering, bias detection systems, and LLM-based autorater frameworks - against these attacks. Results expose significant security vulnerabilities in distilled models, highlighting need for specialized safeguards. We propose practical design principles for building effective adversarial bias mitigation strategies.

Anshuman Suri, Harsh Chaudhari, Yuefeng Peng et al.

While poisoning attacks on machine learning models have been extensively studied, the mechanisms by which adversaries can distribute poisoned models at scale remain largely unexplored. In this paper, we shed light on how model leaderboards -- ranked platforms for model discovery and evaluation -- can serve as a powerful channel for adversaries for stealthy large-scale distribution of poisoned models. We present TrojanClimb, a general framework that enables injection of malicious behaviors while maintaining competitive leaderboard performance. We demonstrate its effectiveness across four diverse modalities: text-embedding, text-generation, text-to-speech and text-to-image, showing that adversaries can successfully achieve high leaderboard rankings while embedding arbitrary harmful functionalities, from backdoors to bias injection. Our findings reveal a significant vulnerability in the machine learning ecosystem, highlighting the urgent need to redesign leaderboard evaluation mechanisms to detect and filter malicious (e.g., poisoned) models, while exposing broader security implications for the machine learning community regarding the risks of adopting models from unverified sources.

Shuyi Lin, Anshuman Suri, Alina Oprea et al.

As large language models (LLMs) become increasingly deployed in safety-critical applications, the lack of systematic methods to assess their vulnerability to jailbreak attacks presents a critical security gap. We introduce the jailbreak oracle problem: given a model, prompt, and decoding strategy, determine whether a jailbreak response can be generated with likelihood exceeding a specified threshold. This formalization enables a principled study of jailbreak vulnerabilities. Answering the jailbreak oracle problem poses significant computational challenges -- the search space grows exponentially with the length of the response tokens. We present Boa, the first efficient algorithm for solving the jailbreak oracle problem. Boa employs a three-phase search strategy: (1) constructing block lists to identify refusal patterns, (2) breadth-first sampling to identify easily accessible jailbreaks, and (3) depth-first priority search guided by fine-grained safety scores to systematically explore promising low-probability paths. Boa enables rigorous security assessments including systematic defense evaluation, standardized comparison of red team attacks, and model certification under extreme adversarial conditions.

Ali Naseh, Harsh Chaudhari, Jaechul Roh et al.

DeepSeek recently released R1, a high-performing large language model (LLM) optimized for reasoning tasks. Despite its efficient training pipeline, R1 achieves competitive performance, even surpassing leading reasoning models like OpenAI's o1 on several benchmarks. However, emerging reports suggest that R1 refuses to answer certain prompts related to politically sensitive topics in China. While existing LLMs often implement safeguards to avoid generating harmful or offensive outputs, R1 represents a notable shift - exhibiting censorship-like behavior on politically charged queries. In this paper, we investigate this phenomenon by first introducing a large-scale set of heavily curated prompts that get censored by R1, covering a range of politically sensitive topics, but are not censored by other models. We then conduct a comprehensive analysis of R1's censorship patterns, examining their consistency, triggers, and variations across topics, prompt phrasing, and context. Beyond English-language queries, we explore censorship behavior in other languages. We also investigate the transferability of censorship to models distilled from the R1 language model. Finally, we propose techniques for bypassing or removing this censorship. Our findings reveal possible additional censorship integration likely shaped by design choices during training or alignment, raising concerns about transparency, bias, and governance in language model deployment.

Georgios Syros, Anshuman Suri, Farinaz Koushanfar et al.

Federated Learning is vulnerable to adversarial manipulation, where malicious clients can inject poisoned updates to influence the global model's behavior. While existing defense mechanisms have made notable progress, they fail to protect against adversaries that aim to induce targeted backdoors under different learning and attack configurations. To address this limitation, we introduce DROP (Distillation-based Reduction Of Poisoning), a novel defense mechanism that combines clustering and activity-tracking techniques with extraction of benign behavior from clients via knowledge distillation to tackle stealthy adversaries that manipulate low data poisoning rates and diverse malicious client ratios within the federation. Through extensive experimentation, our approach demonstrates superior robustness compared to existing defenses across a wide range of learning configurations. Finally, we evaluate existing defenses and our method under the challenging setting of non-IID client data distribution and highlight the challenges of designing a resilient FL defense in this setting.

Ethan Rathbun, Wo Wei Lin, Alina Oprea et al.

Simulated environments are a key piece in the success of Reinforcement Learning (RL), allowing practitioners and researchers to train decision making agents without running expensive experiments on real hardware. Simulators remain a security blind spot, however, enabling adversarial developers to alter the dynamics of their released simulators for malicious purposes. Therefore, in this work we highlight a novel threat, demonstrating how simulator dynamics can be exploited to stealthily implant action-level backdoors into RL agents. The backdoor then allows an adversary to reliably activate targeted actions in an agent upon observing a predefined ``trigger'', leading to potentially dangerous consequences. Traditional backdoor attacks are limited in their strong threat models, assuming the adversary has near full control over an agent's training pipeline, enabling them to both alter and observe agent's rewards. As these assumptions are infeasible to implement within a simulator, we propose a new attack ``Daze'' which is able to reliably and stealthily implant backdoors into RL agents trained for real world tasks without altering or even observing their rewards. We provide formal proof of Daze's effectiveness in guaranteeing attack success across general RL tasks along with extensive empirical evaluations on both discrete and continuous action space domains. We additionally provide the first example of RL backdoor attacks transferring to real, robotic hardware. These developments motivate further research into securing all components of the RL training pipeline to prevent malicious attacks.

Georgios Syros, Evan Rose, Brian Grinstead et al.

Large language model (LLM) based web agents are increasingly deployed to automate complex online tasks by directly interacting with web sites and performing actions on users' behalf. While these agents offer powerful capabilities, their design exposes them to indirect prompt injection attacks embedded in untrusted web content, enabling adversaries to hijack agent behavior and violate user intent. Despite growing awareness of this threat, existing evaluations rely on fixed attack templates, manually selected injection surfaces, or narrowly scoped scenarios, limiting their ability to capture realistic, adaptive attacks encountered in practice. We present MUZZLE, an automated agentic framework for evaluating the security of web agents against indirect prompt injection attacks. MUZZLE utilizes the agent's trajectories to automatically identify high-salience injection surfaces, and adaptively generate context-aware malicious instructions that target violations of confidentiality, integrity, and availability. Unlike prior approaches, MUZZLE adapts its attack strategy based on the agent's observed execution trajectory and iteratively refines attacks using feedback from failed executions. We evaluate MUZZLE across diverse web applications, user tasks, and agent configurations, demonstrating its ability to automatically and adaptively assess the security of web agents with minimal human intervention. Our results show that MUZZLE effectively discovers 37 new attacks on 4 web applications with 10 adversarial objectives that violate confidentiality, availability, or privacy properties. MUZZLE also identifies novel attack strategies, including 2 cross-application prompt injection attacks and an agent-tailored phishing scenario.

Ali Naseh, Anshuman Suri, Yuefeng Peng et al.

Generative AI leaderboards are central to evaluating model capabilities, but remain vulnerable to manipulation. Among key adversarial objectives is rank manipulation, where an attacker must first deanonymize the models behind displayed outputs -- a threat previously demonstrated and explored for large language models (LLMs). We show that this problem can be even more severe for text-to-image leaderboards, where deanonymization is markedly easier. Using over 150,000 generated images from 280 prompts and 19 diverse models spanning multiple organizations, architectures, and sizes, we demonstrate that simple real-time classification in CLIP embedding space identifies the generating model with high accuracy, even without prompt control or historical data. We further introduce a prompt-level separability metric and identify prompts that enable near-perfect deanonymization. Our results indicate that rank manipulation in text-to-image leaderboards is easier than previously recognized, underscoring the need for stronger defenses.

Xavier Cadet, Simona Boboila, Sie Hendrata Dharmawan et al.

Cyber defense requires automating defensive decision-making under stealthy, deceptive, and continuously evolving adversarial strategies. The FlipIt game provides a foundational framework for modeling interactions between a defender and an advanced adversary that compromises a system without being immediately detected. In FlipIt, the attacker and defender compete to control a shared resource by performing a Flip action and paying a cost. However, the existing FlipIt frameworks rely on a small number of heuristics or specialized learning techniques, which can lead to brittleness and the inability to adapt to new attacks. To address these limitations, we introduce PoolFlip, a multi-agent gym environment that extends the FlipIt game to allow efficient learning for attackers and defenders. Furthermore, we propose Flip-PSRO, a multi-agent reinforcement learning (MARL) approach that leverages population-based training to train defender agents equipped to generalize against a range of unknown, potentially adaptive opponents. Our empirical results suggest that Flip-PSRO defenders are $2\times$ more effective than baselines to generalize to a heuristic attack not exposed in training. In addition, our newly designed ownership-based utility functions ensure that Flip-PSRO defenders maintain a high level of control while optimizing performance.

John Abascal, Nicolás Berrios, Alina Oprea et al.

Multitask learning (MTL) has emerged as a powerful paradigm that leverages similarities among multiple learning tasks, each with insufficient samples to train a standalone model, to solve them simultaneously while minimizing data sharing across users and organizations. MTL typically accomplishes this goal by learning a shared representation that captures common structure among the tasks by embedding data from all tasks into a common feature space. Despite being designed to be the smallest unit of shared information necessary to effectively learn patterns across multiple tasks, these shared representations can inadvertently leak sensitive information about the particular tasks they were trained on. In this work, we investigate what information is revealed by the shared representations through the lens of inference attacks. Towards this, we propose a novel, black-box task-inference threat model where the adversary, given the embedding vectors produced by querying the shared representation on samples from a particular task, aims to determine whether that task was present when training the shared representation. We develop efficient, purely black-box attacks on machine learning models that exploit the dependencies between embeddings from the same task without requiring shadow models or labeled reference data. We evaluate our attacks across vision and language domains for multiple use cases of MTL and demonstrate that even with access only to fresh task samples rather than training data, a black-box adversary can successfully infer a task's inclusion in training. To complement our experiments, we provide theoretical analysis of a simplified learning setting and show a strict separation between adversaries with training samples and fresh samples from the target task's distribution.

Xavier Cadet, Simona Boboila, Edward Koh et al.

Cyber resilience is the ability of a system to recover from an attack with minimal impact on system operations. However, characterizing a network's resilience under a cyber attack is challenging, as there are no formal definitions of resilience applicable to diverse network topologies and attack patterns. In this work, we propose a quantifiable formulation of resilience that considers multiple defender operational goals, the criticality of various network resources for daily operations, and provides interpretability to security operators about their system's resilience under attack. We evaluate our approach within the CybORG environment, a reinforcement learning (RL) framework for autonomous cyber defense, analyzing trade-offs between resilience, costs, and prioritization of operational goals. Furthermore, we introduce methods to aggregate resilience metrics across time-variable attack patterns and multiple network topologies, comprehensively characterizing system resilience. Using insights gained from our resilience metrics, we design RL autonomous defensive agents and compare them against several heuristic baselines, showing that proactive network hardening techniques and prompt recovery of compromised machines are critical for effective cyber defenses.

Ethan Rathbun, Alina Oprea, Christopher Amato

Recent works have demonstrated the vulnerability of Deep Reinforcement Learning (DRL) algorithms against training-time, backdoor poisoning attacks. The objectives of these attacks are twofold: induce pre-determined, adversarial behavior in the agent upon observing a fixed trigger during deployment while allowing the agent to solve its intended task during training. Prior attacks assume arbitrary control over the agent's rewards, inducing values far outside the environment's natural constraints. This results in brittle attacks that fail once the proper reward constraints are enforced. Thus, in this work we propose a new class of backdoor attacks against DRL which are the first to achieve state of the art performance under strict reward constraints. These "inception" attacks manipulate the agent's training data -- inserting the trigger into prior observations and replacing high return actions with those of the targeted adversarial behavior. We formally define these attacks and prove they achieve both adversarial objectives against arbitrary Markov Decision Processes (MDP). Using this framework we devise an online inception attack which achieves an 100\% attack success rate on multiple environments under constrained rewards while minimally impacting the agent's task performance.

Andrew Yuan, Alina Oprea, Cheng Tan

Dropout is a common operator in deep learning, aiming to prevent overfitting by randomly dropping neurons during training. This paper introduces a new family of poisoning attacks against neural networks named DROPOUTATTACK. DROPOUTATTACK attacks the dropout operator by manipulating the selection of neurons to drop instead of selecting them uniformly at random. We design, implement, and evaluate four DROPOUTATTACK variants that cover a broad range of scenarios. These attacks can slow or stop training, destroy prediction accuracy of target classes, and sabotage either precision or recall of a target class. In our experiments of training a VGG-16 model on CIFAR-100, our attack can reduce the precision of the victim class by 34.6% (from 81.7% to 47.1%) without incurring any degradation in model accuracy

Krishna Pillutla, Galen Andrew, Peter Kairouz et al.

We present a rigorous methodology for auditing differentially private machine learning algorithms by adding multiple carefully designed examples called canaries. We take a first principles approach based on three key components. First, we introduce Lifted Differential Privacy (LiDP) that expands the definition of differential privacy to handle randomized datasets. This gives us the freedom to design randomized canaries. Second, we audit LiDP by trying to distinguish between the model trained with $K$ canaries versus $K - 1$ canaries in the dataset, leaving one canary out. By drawing the canaries i.i.d., LiDP can leverage the symmetry in the design and reuse each privately trained model to run multiple statistical tests, one for each canary. Third, we introduce novel confidence intervals that take advantage of the multiple test statistics by adapting to the empirical higher-order correlations. Together, this new recipe demonstrates significant improvements in sample complexity, both theoretically and empirically, using synthetic and real data. Further, recent advances in designing stronger canaries can be readily incorporated into the new framework.

Talha Ongun, Oliver Spohngellert, Benjamin Miller et al.

Recent self-propagating malware (SPM) campaigns compromised hundred of thousands of victim machines on the Internet. It is challenging to detect these attacks in their early stages, as adversaries utilize common network services, use novel techniques, and can evade existing detection mechanisms. We propose PORTFILER (PORT-Level Network Traffic ProFILER), a new machine learning system applied to network traffic for detecting SPM attacks. PORTFILER extracts port-level features from the Zeek connection logs collected at a border of a monitored network, applies anomaly detection techniques to identify suspicious events, and ranks the alerts across ports for investigation by the Security Operations Center (SOC). We propose a novel ensemble methodology for aggregating individual models in PORTFILER that increases resilience against several evasion strategies compared to standard ML baselines. We extensively evaluate PORTFILER on traffic collected from two university networks, and show that it can detect SPM attacks with different patterns, such as WannaCry and Mirai, and performs well under evasion. Ranking across ports achieves precision over 0.94 with low false positive rates in the top ranked alerts. When deployed on the university networks, PORTFILER detected anomalous SPM-like activity on one of the campus networks, confirmed by the university SOC as malicious. PORTFILER also detected a Mirai attack recreated on the two university networks with higher precision and recall than deep-learning-based autoencoder methods.

Talha Ongun, Jack W. Stokes, Jonathan Bar Or et al.

In recent years, enterprises have been targeted by advanced adversaries who leverage creative ways to infiltrate their systems and move laterally to gain access to critical data. One increasingly common evasive method is to hide the malicious activity behind a benign program by using tools that are already installed on user computers. These programs are usually part of the operating system distribution or another user-installed binary, therefore this type of attack is called "Living-Off-The-Land". Detecting these attacks is challenging, as adversaries may not create malicious files on the victim computers and anti-virus scans fail to detect them. We propose the design of an Active Learning framework called LOLAL for detecting Living-Off-the-Land attacks that iteratively selects a set of uncertain and anomalous samples for labeling by a human analyst. LOLAL is specifically designed to work well when a limited number of labeled samples are available for training machine learning models to detect attacks. We investigate methods to represent command-line text using word-embedding techniques, and design ensemble boosting classifiers to distinguish malicious and benign samples based on the embedding representation. We leverage a large, anonymized dataset collected by an endpoint security product and demonstrate that our ensemble classifiers achieve an average F1 score of 0.96 at classifying different attack classes. We show that our active learning method consistently improves the classifier performance, as more training data is labeled, and converges in less than 30 iterations when starting with a small number of labeled instances.

Lisa Oakley, Alina Oprea, Stavros Tripakis

Probabilistic model checking is a useful technique for specifying and verifying properties of stochastic systems including randomized protocols and reinforcement learning models. Existing methods rely on the assumed structure and probabilities of certain system transitions. These assumptions may be incorrect, and may even be violated by an adversary who gains control of system components. In this paper, we develop a formal framework for adversarial robustness in systems modeled as discrete time Markov chains (DTMCs). We base our framework on existing methods for verifying probabilistic temporal logic properties and extend it to include deterministic, memoryless policies acting in Markov decision processes (MDPs). Our framework includes a flexible approach for specifying structure-preserving and non structure-preserving adversarial models. We outline a class of threat models under which adversaries can perturb system transitions, constrained by an $\varepsilon$ ball around the original transition probabilities. We define three main DTMC adversarial robustness problems: adversarial robustness verification, maximal $δ$ synthesis, and worst case attack synthesis. We present two optimization-based solutions to these three problems, leveraging traditional and parametric probabilistic model checking techniques. We then evaluate our solutions on two stochastic protocols and a collection of Grid World case studies, which model an agent acting in an environment described as an MDP. We find that the parametric solution results in fast computation for small parameter spaces. In the case of less restrictive (stronger) adversaries, the number of parameters increases, and directly computing property satisfaction probabilities is more scalable. We demonstrate the usefulness of our definitions and solutions by comparing system outcomes over various properties, threat models, and case studies.

Talha Ongun, Simona Boboila, Alina Oprea et al.

Recently, coordinated attack campaigns started to become more widespread on the Internet. In May 2017, WannaCry infected more than 300,000 machines in 150 countries in a few days and had a large impact on critical infrastructure. Existing threat sharing platforms cannot easily adapt to emerging attack patterns. At the same time, enterprises started to adopt machine learning-based threat detection tools in their local networks. In this paper, we pose the question: \emph{What information can defenders share across multiple networks to help machine learning-based threat detection adapt to new coordinated attacks?} We propose three information sharing methods across two networks, and show how the shared information can be used in a machine-learning network-traffic model to significantly improve its ability of detecting evasive self-propagating malware.

Molly Buchanan, Jeffrey W. Collyer, Jack W. Davidson et al.

Research and development of techniques which detect or remediate malicious network activity require access to diverse, realistic, contemporary data sets containing labeled malicious connections. In the absence of such data, said techniques cannot be meaningfully trained, tested, and evaluated. Synthetically produced data containing fabricated or merged network traffic is of limited value as it is easily distinguishable from real traffic by even simple machine learning (ML) algorithms. Real network data is preferable, but while ubiquitous is broadly both sensitive and lacking in ground truth labels, limiting its utility for ML research. This paper presents a multi-faceted approach to generating a data set of labeled malicious connections embedded within anonymized network traffic collected from large production networks. Real-world malware is defanged and introduced to simulated, secured nodes within those networks to generate realistic traffic while maintaining sufficient isolation to protect real data and infrastructure. Network sensor data, including this embedded malware traffic, is collected at a network edge and anonymized for research use. Network traffic was collected and produced in accordance with the aforementioned methods at two major educational institutions. The result is a highly realistic, long term, multi-institution data set with embedded data labels spanning over 1.5 trillion connections and over a petabyte of sensor log data. The usability of this data set is demonstrated by its utility to our artificial intelligence and machine learning (AI/ML) research program.

Nicholas Carlini, Florian Tramer, Eric Wallace et al.

It has become common to publish large (billion parameter) language models that have been trained on private datasets. This paper demonstrates that in such settings, an adversary can perform a training data extraction attack to recover individual training examples by querying the language model. We demonstrate our attack on GPT-2, a language model trained on scrapes of the public Internet, and are able to extract hundreds of verbatim text sequences from the model's training data. These extracted examples include (public) personally identifiable information (names, phone numbers, and email addresses), IRC conversations, code, and 128-bit UUIDs. Our attack is possible even though each of the above sequences are included in just one document in the training data. We comprehensively evaluate our extraction attack to understand the factors that contribute to its success. Worryingly, we find that larger models are more vulnerable than smaller models. We conclude by drawing lessons and discussing possible safeguards for training large language models.