David Wagner

Norman Mu, Sarah Chen, Zifan Wang et al. · berkeley

As Large Language Models (LLMs) are deployed with increasing real-world responsibilities, it is important to be able to specify and constrain the behavior of these systems in a reliable manner. Model developers may wish to set explicit rules for the model, such as "do not generate abusive content", but these may be circumvented by jailbreaking techniques. Existing evaluations of adversarial attacks and defenses on LLMs generally require either expensive manual review or unreliable heuristic checks. To address this issue, we propose Rule-following Language Evaluation Scenarios (RuLES), a programmatic framework for measuring rule-following ability in LLMs. RuLES consists of 14 simple text scenarios in which the model is instructed to obey various rules while interacting with the user. Each scenario has a programmatic evaluation function to determine whether the model has broken any rules in a conversation. Our evaluations of proprietary and open models show that almost all current models struggle to follow scenario rules, even on straightforward test cases. We also demonstrate that simple optimization attacks suffice to significantly increase failure rates on test cases. We conclude by exploring two potential avenues for improvement: test-time steering and supervised fine-tuning.

Nabeel Hingun, Chawin Sitawarin, Jerry Li et al.

Machine learning models are known to be susceptible to adversarial perturbation. One famous attack is the adversarial patch, a sticker with a particularly crafted pattern that makes the model incorrectly predict the object it is placed on. This attack presents a critical threat to cyber-physical systems that rely on cameras such as autonomous cars. Despite the significance of the problem, conducting research in this setting has been difficult; evaluating attacks and defenses in the real world is exceptionally costly while synthetic data are unrealistic. In this work, we propose the REAP (REalistic Adversarial Patch) benchmark, a digital benchmark that allows the user to evaluate patch attacks on real images, and under real-world conditions. Built on top of the Mapillary Vistas dataset, our benchmark contains over 14,000 traffic signs. Each sign is augmented with a pair of geometric and lighting transformations, which can be used to apply a digitally generated patch realistically onto the sign. Using our benchmark, we perform the first large-scale assessments of adversarial patch attacks under realistic conditions. Our experiments suggest that adversarial patch attacks may present a smaller threat than previously believed and that the success rate of an attack on simpler digital simulations is not predictive of its actual effectiveness in practice. We release our benchmark publicly at https://github.com/wagner-group/reap-benchmark.

Chawin Sitawarin, Zachary Golan-Strieb, David Wagner

Neural networks' lack of robustness against attacks raises concerns in security-sensitive settings such as autonomous vehicles. While many countermeasures may look promising, only a few withstand rigorous evaluation. Defenses using random transformations (RT) have shown impressive results, particularly BaRT (Raff et al., 2019) on ImageNet. However, this type of defense has not been rigorously evaluated, leaving its robustness properties poorly understood. Their stochastic properties make evaluation more challenging and render many proposed attacks on deterministic models inapplicable. First, we show that the BPDA attack (Athalye et al., 2018a) used in BaRT's evaluation is ineffective and likely overestimates its robustness. We then attempt to construct the strongest possible RT defense through the informed selection of transformations and Bayesian optimization for tuning their parameters. Furthermore, we create the strongest possible attack to evaluate our RT defense. Our new attack vastly outperforms the baseline, reducing the accuracy by 83% compared to the 19% reduction by the commonly used EoT attack ($4.3\times$ improvement). Our result indicates that the RT defense on the Imagenette dataset (a ten-class subset of ImageNet) is not robust against adversarial examples. Extending the study further, we use our new attack to adversarially train RT defense (called AdvRT), resulting in a large robustness gain. Code is available at https://github.com/wagner-group/demystify-random-transform.

Chawin Sitawarin, Kornrapat Pongmala, Yizheng Chen et al.

We show that combining human prior knowledge with end-to-end learning can improve the robustness of deep neural networks by introducing a part-based model for object classification. We believe that the richer form of annotation helps guide neural networks to learn more robust features without requiring more samples or larger models. Our model combines a part segmentation model with a tiny classifier and is trained end-to-end to simultaneously segment objects into parts and then classify the segmented object. Empirically, our part-based models achieve both higher accuracy and higher adversarial robustness than a ResNet-50 baseline on all three datasets. For instance, the clean accuracy of our part models is up to 15 percentage points higher than the baseline's, given the same level of robustness. Our experiments indicate that these models also reduce texture bias and yield better robustness against common corruptions and spurious correlations. The code is publicly available at https://github.com/chawins/adv-part-model.

Yizheng Chen, Zhoujie Ding, Lamya Alowain et al.

We propose and release a new vulnerable source code dataset. We curate the dataset by crawling security issue websites, extracting vulnerability-fixing commits and source codes from the corresponding projects. Our new dataset contains 18,945 vulnerable functions spanning 150 CWEs and 330,492 non-vulnerable functions extracted from 7,514 commits. Our dataset covers 295 more projects than all previous datasets combined. Combining our new dataset with previous datasets, we present an analysis of the challenges and promising research directions of using deep learning for detecting software vulnerabilities. We study 11 model architectures belonging to 4 families. Our results show that deep learning is still not ready for vulnerability detection, due to high false positive rate, low F1 score, and difficulty of detecting hard CWEs. In particular, we demonstrate an important generalization challenge for the deployment of deep learning-based models. We show that increasing the volume of training data may not further improve the performance of deep learning models for vulnerability detection, but might be useful to improve the generalization ability to unseen projects. We also identify hopeful future research directions. We demonstrate that large language models (LLMs) are a promising research direction for ML-based vulnerability detection, outperforming Graph Neural Networks (GNNs) with code-structure features in our experiments. Moreover, developing source code specific pre-training objectives is a promising research direction to improve the vulnerability detection performance.

Chawin Sitawarin, Jaewon Chang, David Huang et al.

Adversarial attacks have been a looming and unaddressed threat in the industry. However, through a decade-long history of the robustness evaluation literature, we have learned that mounting a strong or optimal attack is challenging. It requires both machine learning and domain expertise. In other words, the white-box threat model, religiously assumed by a large majority of the past literature, is unrealistic. In this paper, we propose a new practical threat model where the adversary relies on transfer attacks through publicly available surrogate models. We argue that this setting will become the most prevalent for security-sensitive applications in the future. We evaluate the transfer attacks in this setting and propose a specialized defense method based on a game-theoretic perspective. The defenses are evaluated under 24 public models and 11 attack algorithms across three datasets (CIFAR-10, CIFAR-100, and ImageNet). Under this threat model, our defense, PubDef, outperforms the state-of-the-art white-box adversarial training by a large margin with almost no loss in the normal accuracy. For instance, on ImageNet, our defense achieves 62% accuracy under the strongest transfer attack vs only 36% of the best adversarially trained model. Its accuracy when not under attack is only 2% lower than that of an undefended model (78% vs 80%). We release our code at https://github.com/wagner-group/pubdef.

Yizheng Chen, Zhoujie Ding, David Wagner

Machine learning methods can detect Android malware with very high accuracy. However, these classifiers have an Achilles heel, concept drift: they rapidly become out of date and ineffective, due to the evolution of malware apps and benign apps. Our research finds that, after training an Android malware classifier on one year's worth of data, the F1 score quickly dropped from 0.99 to 0.76 after 6 months of deployment on new test samples. In this paper, we propose new methods to combat the concept drift problem of Android malware classifiers. Since machine learning technique needs to be continuously deployed, we use active learning: we select new samples for analysts to label, and then add the labeled samples to the training set to retrain the classifier. Our key idea is, similarity-based uncertainty is more robust against concept drift. Therefore, we combine contrastive learning with active learning. We propose a new hierarchical contrastive learning scheme, and a new sample selection technique to continuously train the Android malware classifier. Our evaluation shows that this leads to significant improvements, compared to previously published methods for active learning. Our approach reduces the false negative rate from 14% (for the best baseline) to 9%, while also reducing the false positive rate (from 0.86% to 0.48%). Also, our approach maintains more consistent performance across a seven-year time period than past methods.

Julien Piet, Maha Alrashed, Chawin Sitawarin et al. · pku

Large Language Models (LLMs) are attracting significant research attention due to their instruction-following abilities, allowing users and developers to leverage LLMs for a variety of tasks. However, LLMs are vulnerable to prompt-injection attacks: a class of attacks that hijack the model's instruction-following abilities, changing responses to prompts to undesired, possibly malicious ones. In this work, we introduce Jatmo, a method for generating task-specific models resilient to prompt-injection attacks. Jatmo leverages the fact that LLMs can only follow instructions once they have undergone instruction tuning. It harnesses a teacher instruction-tuned model to generate a task-specific dataset, which is then used to fine-tune a base model (i.e., a non-instruction-tuned model). Jatmo only needs a task prompt and a dataset of inputs for the task: it uses the teacher model to generate outputs. For situations with no pre-existing datasets, Jatmo can use a single example, or in some cases none at all, to produce a fully synthetic dataset. Our experiments on seven tasks show that Jatmo models provide similar quality of outputs on their specific task as standard LLMs, while being resilient to prompt injections. The best attacks succeeded in less than 0.5% of cases against our models, versus 87% success rate against GPT-3.5-Turbo. We release Jatmo at https://github.com/wagner-group/prompt-injection-defense.

Evan Luo, Julien Piet, David Wagner

Security analysts routinely query system logs to detect threats and investigate incidents, but each log source uses its own semi-structured format: logs are cheap to produce, but expensive to use. The standard approach, building per-source parsers to normalize logs into structured schemas, is powerful but requires continuous engineering effort for each new format. Querying raw logs directly with tools like grep avoids this cost, but requires analysts to know each source's message variants and cannot express the multi-line temporal queries that security investigations demand. We present Sieve, a system that generates executable query code from natural-language security questions by grounding a large language model with lightweight, automatically extracted log-format context, requiring only one LLM call per query followed by deterministic execution. Evaluating 133 security queries across 5 log types, we find that Sieve achieves over a 3x reduction in error rate on complex temporal and cross-event queries compared to manual analyst scripting, with the largest gains on the multi-line correlation tasks most critical to active investigations. Our results and benchmark provide evidence that LLM-generated code can bridge the gap between the expressiveness of structured log querying and the immediacy of working directly with raw files.

Chawin Sitawarin, Norman Mu, David Wagner et al.

Large Language Models (LLMs) have surged in popularity in recent months, but they have demonstrated concerning capabilities to generate harmful content when manipulated. While techniques like safety fine-tuning aim to minimize harmful use, recent works have shown that LLMs remain vulnerable to attacks that elicit toxic responses. In this work, we introduce the Proxy-Guided Attack on LLMs (PAL), the first optimization-based attack on LLMs in a black-box query-only setting. In particular, it relies on a surrogate model to guide the optimization and a sophisticated loss designed for real-world LLM APIs. Our attack achieves 84% attack success rate (ASR) on GPT-3.5-Turbo and 48% on Llama-2-7B, compared to 4% for the current state of the art. We also propose GCG++, an improvement to the GCG attack that reaches 94% ASR on white-box Llama-2-7B, and the Random-Search Attack on LLMs (RAL), a strong but simple baseline for query-based attacks. We believe the techniques proposed in this work will enable more comprehensive safety testing of LLMs and, in the long term, the development of better security guardrails. The code can be found at https://github.com/chawins/pal.

Sizhe Chen, Arman Zharmagambetov, David Wagner et al.

Prompt injection attack has been listed as the top-1 security threat to LLM-integrated applications, which interact with external environment data for complex tasks. The untrusted data may contain an injected prompt trying to arbitrarily manipulate the system. Model-level prompt injection defenses have shown strong effectiveness, but are currently deployed into commercial-grade models in a closed-source manner. We believe open-source secure models are needed by the AI security community, where co-development of attacks and defenses through open research drives scientific progress in mitigating prompt injection attacks. To this end, we develop Meta SecAlign, the first fully open-source LLM with built-in model-level defense that achieves commercial-grade performance, powerful enough for complex agentic tasks. We provide complete details of our training recipe, an improved version of the SOTA SecAlign defense. We perform the most comprehensive evaluation to date on 9 utility benchmarks and 7 security benchmarks on general knowledge, instruction following, and agentic workflows. Results show that Meta SecAlign, despite being trained on generic instruction-tuning samples only, surprisingly confers security in unseen downstream tasks, including tool-calling and web-navigation, in addition to general instruction-following. Our best model -- Meta-SecAlign-70B -- establishes a new frontier of utility-security trade-off for open-source LLMs. Even compared to closed-course commercial models such as GPT-5, our model is much securer than most of them. Below are links for the code (https://github.com/facebookresearch/Meta_SecAlign), Meta-SecAlign-70B(https://huggingface.co/facebook/Meta-SecAlign-70B), and Meta-SecAlign-8B(https://huggingface.co/facebook/Meta-SecAlign-8B) models.

Julien Piet, Annabella Chow, Yiwei Hou et al.

ReAct has become the default architecture across LLM agents, and many existing web agents follow this paradigm. We argue that it is the wrong default for web agents. Instead, web agents should default to plan-then-execute: commit to a task-specific program before observing runtime web content, then execute it. The reason is that web content mixes inputs from many parties. An e-commerce product page may combine a seller's listing, customer reviews and sponsored advertisements. Under ReAct, all of this content flows into the model when deciding on the next action, creating a direct path for prompt injections to steer the agent's control flow. Plan-then-execute changes this boundary: untrusted data may influence values or branches inside a predefined execution graph, but it cannot redefine the user task or cause the model to synthesize new actions at runtime. We analyze WebArena, a popular web agent benchmark, and find that all tasks are compatible with plan-then-execute, while 80% can be completed with a purely programmatic plan, without any runtime LLM subroutine. We identify the main barrier to adopting plan-then-execute on the web: For it to work well, tools must map cleanly to semantic actions, with effects known before execution, so agents have enough information to plan. The web does not naturally expose that interface. Browser tools such as click, type, and scroll have page-dependent meanings. Planning at this layer is near-sighted: the agent can only see actions on the current page, and later actions appear only after it acts. Closing this gap requires typed interfaces that turn website interactions from clicks and keystrokes to task-level operations. This is an infrastructure problem, not a modeling problem. Web tasks do not need reactivity by default; they need typed, complete, auditable website APIs.

Zhanhao Hu, Xiao Huang, Patrick Mendoza et al.

Large Language Models (LLMs) pose a significant risk of safety misalignment after finetuning, as models can be compromised by both explicitly and implicitly harmful data. Even some seemingly benign data can inadvertently steer a model towards misaligned behaviors. To address this, we introduce GradShield, a principled filtering method that safeguards LLMs during finetuning by identifying and removing harmful data points before they corrupt the model's alignment. It removes potentially harmful data by computing a Finetuning Implicit Harmfulness Score (FIHS) for each data point and employs an adaptive thresholding algorithm. We apply GradShield to multiple utility fine-tuning tasks across varying levels of harmful data and evaluate the safety and utility performance of the resulting LLMs using various metrics. The results show that GradShield outperforms all baseline methods, consistently maintaining an Attack Success Rate (ASR) below $6\%$ while preserving utility performance.

Yangruibo Ding, Yanjun Fu, Omniyyah Ibrahim et al.

In the context of the rising interest in code language models (code LMs) and vulnerability detection, we study the effectiveness of code LMs for detecting vulnerabilities. Our analysis reveals significant shortcomings in existing vulnerability datasets, including poor data quality, low label accuracy, and high duplication rates, leading to unreliable model performance in realistic vulnerability detection scenarios. Additionally, the evaluation methods used with these datasets are not representative of real-world vulnerability detection. To address these challenges, we introduce PrimeVul, a new dataset for training and evaluating code LMs for vulnerability detection. PrimeVul incorporates a novel set of data labeling techniques that achieve comparable label accuracy to human-verified benchmarks while significantly expanding the dataset. It also implements a rigorous data de-duplication and chronological data splitting strategy to mitigate data leakage issues, alongside introducing more realistic evaluation metrics and settings. This comprehensive approach aims to provide a more accurate assessment of code LMs' performance in real-world conditions. Evaluating code LMs on PrimeVul reveals that existing benchmarks significantly overestimate the performance of these models. For instance, a state-of-the-art 7B model scored 68.26% F1 on BigVul but only 3.09% F1 on PrimeVul. Attempts to improve performance through advanced training techniques and larger models like GPT-3.5 and GPT-4 were unsuccessful, with results akin to random guessing in the most stringent settings. These findings underscore the considerable gap between current capabilities and the practical requirements for deploying code LMs in security roles, highlighting the need for more innovative research in this domain.

Chong Xiang, Tong Wu, Zexuan Zhong et al. · princeton

Retrieval-augmented generation (RAG) has been shown vulnerable to retrieval corruption attacks: an attacker can inject malicious passages into retrieval results to induce inaccurate responses. In this paper, we propose RobustRAG as the first defense framework against retrieval corruption attacks. The key insight of RobustRAG is an isolate-then-aggregate strategy: we get LLM responses from each passage in isolation and then securely aggregate these isolated responses. To instantiate RobustRAG, we design keyword-based and decoding-based algorithms for securely aggregating unstructured text responses. Notably, RobustRAG can achieve certifiable robustness: we can formally prove and certify that, for certain queries, RobustRAG can always return accurate responses, even when the attacker has full knowledge of our defense and can arbitrarily inject a small number of malicious passages. We evaluate RobustRAG on open-domain QA and long-form text generation datasets and demonstrate its effectiveness and generalizability across various tasks and datasets.

Debeshee Das, Julien Piet, Darya Kaviani et al.

Memory systems enable otherwise-stateless LLM agents to persist user information across sessions, but also introduce a new attack surface. We characterize the Trojan Hippo attack, a class of persistent memory attacks that operates in a more realistic threat model than prior memory poisoning work: the attacker plants a dormant payload into an agent's long-term memory via a single untrusted tool call (e.g., a crafted email), which activates only when the user later discusses sensitive topics such as finance, health, or identity, and exfiltrates high-value personal data to the attacker. While anecdotal demonstrations of such attacks have appeared against deployed systems, no prior work systematically evaluates them across heterogeneous memory architectures and defenses.We introduce a dynamic evaluation framework comprising two components: (1) an OpenEvolve-based adaptive red-teaming benchmark that stress-tests defenses and memory backends against continuously refined attacks, and (2) the first capability-aware security/utility analysis for persistent memory systems, enabling principled reasoning about defense deployment across different usage profiles. Instantiated on an email assistant across four memory backends (explicit tool memory, agentic memory, RAG, and sliding-window context), Trojan Hippo achieves up to 85-100 percent ASR against current frontier models from OpenAI and Google, with planted memories successfully activating even after 100 benign sessions. We evaluate four memory-system defenses inspired by basic security principles, finding they substantially reduce attack success rates (to as low as 0-5 percent), though at utility costs that vary widely with task requirements. Because of this substantial security-utility tradeoff, the effective real-world deployment of defenses remains an open challenge, which our evaluation framework is specifically designed to address.

Rebecca Saul, Jingzhi Jiang, Elliott Chia et al.

Traditionally, machine learning methods for PE malware detection have relied on static features like byte histograms, string information, and PE header contents. One barrier to incorporating dynamic analysis features has been the semi-structured nature of sandbox behavior reports. We show that, using the latest generation of large language models with reasoning, it is possible to efficiently process these behavior reports and utilize them as part of a malware detection pipeline. Specifically, we leverage LLMs to generate behavior-based malware detection rules based on a small training set of labeled malware. We find that these detection rules, derived from behavioral features, are much more robust to concept drift than standard static-feature methods, while maintaining practical false positive rates. Finally, we introduce Trident, a system which combines a classic decision tree model over static features, our behavior-based detection rules, and direct LLM analysis of sandbox reports through majority voting. Trident outperforms standard methods using static features, outperforms behavior-based rules alone, and is as resilient to concept drift as active learning methods without requiring retraining.

Ricky Sahu, Eric Marriott, Ethan Siegel et al.

With U.S. healthcare spending approaching $5T (NHE Fact Sheet 2024), and 25% of it estimated to be wasteful (Waste in the US the health care system: estimated costs and potential for savings, n.d.), the need to better predict risk and optimal patient care is evermore important. This paper introduces the Large Medical Model (LMM), a generative pre-trained transformer (GPT) designed to guide and predict the broad facets of patient care and healthcare administration. The model is trained on medical event sequences from over 140M longitudinal patient claims records with a specialized vocabulary built from medical terminology systems and demonstrates a superior capability to forecast healthcare costs and identify potential risk factors. Through experimentation and validation, we showcase the LMM's proficiency in not only in cost and risk predictions, but also in discerning intricate patterns within complex medical conditions and an ability to identify novel relationships in patient care. The LMM is able to improve both cost prediction by 14.1% over the best commercial models and chronic conditions prediction by 1.9% over the best transformer models in research predicting a broad set of conditions. The LMM is a substantial advancement in healthcare analytics, offering the potential to significantly enhance risk assessment, cost management, and personalized medicine.

Banghua Zhu, Norman Mu, Jiantao Jiao et al.

Generative AI's expanding footprint across numerous industries has led to both excitement and increased scrutiny. This paper delves into the unique security challenges posed by Generative AI, and outlines potential research directions for managing these risks.

Norman Mu, Jonathan Lu, Michael Lavery et al.

System prompts have emerged as a critical control surface for specifying the behavior of LLMs in chat and agent settings. Developers depend on system prompts to specify important context, output format, personalities, guardrails, content policies, and safety countermeasures, all of which require models to robustly adhere to the system prompt, especially when facing conflicting or adversarial user inputs. In practice, models often forget to consider relevant guardrails or fail to resolve conflicting demands between the system and the user. In this work, we study various methods for improving system prompt robustness by creating realistic new evaluation and fine-tuning datasets based on prompts collected from from OpenAI's GPT Store and HuggingFace's HuggingChat. Our experiments assessing models with a panel of new and existing benchmarks show that performance can be considerably improved with realistic fine-tuning data, as well as inference-time interventions such as classifier-free guidance. Finally, we analyze the results of recently released reasoning models from OpenAI and DeepSeek, which show exciting but uneven improvements on the benchmarks we study. Overall, current techniques fall short of ensuring system prompt robustness and further study is warranted.

Jesson Wang, Zhanhao Hu, David Wagner

Large Language Models (LLMs) are trained with safety alignment to prevent generating malicious content. Although some attacks have highlighted vulnerabilities in these safety-aligned LLMs, they typically have limitations, such as necessitating access to the model weights or the generation process. Since proprietary models through API-calling do not grant users such permissions, these attacks find it challenging to compromise them. In this paper, we propose Jailbreaking Using LLM Introspection (JULI), which jailbreaks LLMs by manipulating the token log probabilities, using a tiny plug-in block, BiasNet. JULI relies solely on the knowledge of the target LLM's predicted token log probabilities. It can effectively jailbreak API-calling LLMs under a black-box setting and knowing only top-$5$ token log probabilities. Our approach demonstrates superior effectiveness, outperforming existing state-of-the-art (SOTA) approaches across multiple metrics.

Aidan Wen, Norah A. Alzahrani, Jingzhi Jiang et al.

We introduce SeedAIchemy, an automated LLM-driven corpus generation tool that makes it easier for developers to implement fuzzing effectively. SeedAIchemy consists of five modules which implement different approaches at collecting publicly available files from the internet. Four of the five modules use large language model (LLM) workflows to construct search terms designed to maximize corpus quality. Corpora generated by SeedAIchemy perform significantly better than a naive corpus and similarly to a manually-curated corpus on a diverse range of target programs and libraries.

Dennis Jacob, Emad Alghamdi, Zhanhao Hu et al.

Large language models (LLMs) have become increasingly popular due to their ability to interact with unstructured content. As such, LLMs are now a key driver behind the automation of language processing systems, such as AI agents. Unfortunately, these advantages have come with a vulnerability to prompt injections, an attack where an adversary subverts the LLM's intended functionality with an injected task. Past approaches have proposed detectors and finetuning to provide robustness, but these techniques are vulnerable to adaptive attacks or cannot be used with state-of-the-art models. To this end we propose type-directed privilege separation for LLMs, a method that systematically prevents prompt injections. We restrict the ability of an LLM to interact with third-party data by converting untrusted content to a curated set of data types; unlike raw strings, each data type is limited in scope and content, eliminating the possibility for prompt injections. We evaluate our method across several case studies and find that designs leveraging our principles can systematically prevent prompt injection attacks while maintaining high utility.

Paul Francis, Sebastian Probst-Eide, David Wagner et al.

Historically, strong data anonymization requires substantial domain expertise and custom design for the given data set and use case. Diffix is an anonymization framework designed to make strong data anonymization available to non-experts. This paper describes Diffix Elm, a version of Diffix that is very easy to use at the expense of query features. We describe Diffix Elm, and show that it provides strong anonymity based on the General Data Protection Regulation (GDPR) criteria. This document is the third version of Diffix Elm. The second version added ceiling, round, and bucket\_width functions (in addition to floor). This document adds the ability to protect multiple different kinds of protected entities (a feature not found in earlier versions of Diffix). It also adds counting distinct values for any column (rather than only the AID column).

Norman Mu, Alexander Kirillov, David Wagner et al.

Recent work has shown that self-supervised pre-training leads to improvements over supervised learning on challenging visual recognition tasks. CLIP, an exciting new approach to learning with language supervision, demonstrates promising performance on a wide variety of benchmarks. In this work, we explore whether self-supervised learning can aid in the use of language supervision for visual representation learning. We introduce SLIP, a multi-task learning framework for combining self-supervised learning and CLIP pre-training. After pre-training with Vision Transformers, we thoroughly evaluate representation quality and compare performance to both CLIP and self-supervised learning under three distinct settings: zero-shot transfer, linear classification, and end-to-end finetuning. Across ImageNet and a battery of additional datasets, we find that SLIP improves accuracy by a large margin. We validate our results further with experiments on different model sizes, training schedules, and pre-training datasets. Our findings show that SLIP enjoys the best of both worlds: better performance than self-supervision (+8.1% linear accuracy) and language supervision (+5.2% zero-shot accuracy).

Grant Ho, Mayank Dhiman, Devdatta Akhawe et al.

In successful enterprise attacks, adversaries often need to gain access to additional machines beyond their initial point of compromise, a set of internal movements known as lateral movement. We present Hopper, a system for detecting lateral movement based on commonly available enterprise logs. Hopper constructs a graph of login activity among internal machines and then identifies suspicious sequences of loginsthat correspond to lateral movement. To understand the larger context of each login, Hopper employs an inference algorithm to identify the broader path(s) of movement that each login belongs to and the causal user responsible for performing a path's logins. Hopper then leverages this path inference algorithm, in conjunction with a set of detection rules and a new anomaly scoring algorithm, to surface the login paths most likely to reflect lateral movement. On a 15-month enterprise dataset consisting of over 780 million internal logins, Hop-per achieves a 94.5% detection rate across over 300 realistic attack scenarios, including one red team attack, while generating an average of <9 alerts per day. In contrast, to detect the same number of attacks, prior state-of-the-art systems would need to generate nearly 8x as many false positives.

Yizheng Chen, Shiqi Wang, Yue Qin et al.

Many recent works have proposed methods to train classifiers with local robustness properties, which can provably eliminate classes of evasion attacks for most inputs, but not all inputs. Since data distribution shift is very common in security applications, e.g., often observed for malware detection, local robustness cannot guarantee that the property holds for unseen inputs at the time of deploying the classifier. Therefore, it is more desirable to enforce global robustness properties that hold for all inputs, which is strictly stronger than local robustness. In this paper, we present a framework and tools for training classifiers that satisfy global robustness properties. We define new notions of global robustness that are more suitable for security classifiers. We design a novel booster-fixer training framework to enforce global robustness properties. We structure our classifier as an ensemble of logic rules and design a new verifier to verify the properties. In our training algorithm, the booster increases the classifier's capacity, and the fixer enforces verified global robustness properties following counterexample guided inductive synthesis. We show that we can train classifiers to satisfy different global robustness properties for three security datasets, and even multiple properties at the same time, with modest impact on the classifier's performance. For example, we train a Twitter spam account classifier to satisfy five global robustness properties, with 5.4% decrease in true positive rate, and 0.1% increase in false positive rate, compared to a baseline XGBoost model that doesn't satisfy any property.

Dequan Wang, An Ju, Evan Shelhamer et al.

Adversarial attacks optimize against models to defeat defenses. Existing defenses are static, and stay the same once trained, even while attacks change. We argue that models should fight back, and optimize their defenses against attacks at test time. We propose dynamic defenses, to adapt the model and input during testing, by defensive entropy minimization (dent). Dent alters testing, but not training, for compatibility with existing models and train-time defenses. Dent improves the robustness of adversarially-trained defenses and nominally-trained models against white-box, black-box, and adaptive attacks on CIFAR-10/100 and ImageNet. In particular, dent boosts state-of-the-art defenses by 20+ points absolute against AutoAttack on CIFAR-10 at $ε_\infty$ = 8/255.

Henry Xu, An Ju, David Wagner

Susceptibility of neural networks to adversarial attack prompts serious safety concerns for lane detection efforts, a domain where such models have been widely applied. Recent work on adversarial road patches have successfully induced perception of lane lines with arbitrary form, presenting an avenue for rogue control of vehicle behavior. In this paper, we propose a modular lane verification system that can catch such threats before the autonomous driving system is misled while remaining agnostic to the particular lane detection model. Our experiments show that implementing the system with a simple convolutional neural network (CNN) can defend against a wide gamut of attacks on lane detection models. With a 10% impact to inference time, we can detect 96% of bounded non-adaptive attacks, 90% of bounded adaptive attacks, and 98% of patch attacks while preserving accurate identification at least 95% of true lanes, indicating that our proposed verification system is effective at mitigating lane detection security risks with minimal overhead.

Chawin Sitawarin, Evgenios M. Kornaropoulos, Dawn Song et al.

Adversarial examples are a widely studied phenomenon in machine learning models. While most of the attention has been focused on neural networks, other practical models also suffer from this issue. In this work, we propose an algorithm for evaluating the adversarial robustness of $k$-nearest neighbor classification, i.e., finding a minimum-norm adversarial example. Diverging from previous proposals, we take a geometric approach by performing a search that expands outwards from a given input point. On a high level, the search radius expands to the nearby Voronoi cells until we find a cell that classifies differently from the input point. To scale the algorithm to a large $k$, we introduce approximation steps that find perturbations with smaller norm, compared to the baselines, in a variety of datasets. Furthermore, we analyze the structural properties of a dataset where our approach outperforms the competition.

Neil Shah, Grant Ho, Marco Schweighauser et al.

We present a large-scale characterization of attacker activity across 111 real-world enterprise organizations. We develop a novel forensic technique for distinguishing between attacker activity and benign activity in compromised enterprise accounts that yields few false positives and enables us to perform fine-grained analysis of attacker behavior. Applying our methods to a set of 159 compromised enterprise accounts, we quantify the duration of time attackers are active in accounts and examine thematic patterns in how attackers access and leverage these hijacked accounts. We find that attackers frequently dwell in accounts for multiple days to weeks, suggesting that delayed (non-real-time) detection can still provide significant value. Based on an analysis of the attackers' timing patterns, we observe two distinct modalities in how attackers access compromised accounts, which could be explained by the existence of a specialized market for hijacked enterprise accounts: where one class of attackers focuses on compromising and selling account access to another class of attackers who exploit the access such hijacked accounts provide. Ultimately, our analysis sheds light on the state of enterprise account hijacking and highlights fruitful directions for a broader space of detection methods, ranging from new features that home in on malicious account behavior to the development of non-real-time detection methods that leverage malicious activity after an attack's initial point of compromise to more accurately identify attacks.

Michael McCoyd, Won Park, Steven Chen et al.

Deep learning image classification is vulnerable to adversarial attack, even if the attacker changes just a small patch of the image. We propose a defense against patch attacks based on partially occluding the image around each candidate patch location, so that a few occlusions each completely hide the patch. We demonstrate on CIFAR-10, Fashion MNIST, and MNIST that our defense provides certified security against patch attacks of a certain size.

Chawin Sitawarin, Supriyo Chakraborty, David Wagner

Adversarial training (AT) has become a popular choice for training robust networks. However, it tends to sacrifice clean accuracy heavily in favor of robustness and suffers from a large generalization error. To address these concerns, we propose Smooth Adversarial Training (SAT), guided by our analysis on the eigenspectrum of the loss Hessian. We find that curriculum learning, a scheme that emphasizes on starting "easy" and gradually ramping up on the "difficulty" of training, smooths the adversarial loss landscape for a suitably chosen difficulty metric. We present a general formulation for curriculum learning in the adversarial setting and propose two difficulty metrics based on the maximal Hessian eigenvalue (H-SAT) and the softmax probability (P-SA). We demonstrate that SAT stabilizes network training even for a large perturbation norm and allows the network to operate at a better clean accuracy versus robustness trade-off curve compared to AT. This leads to a significant improvement in both clean accuracy and robustness compared to AT, TRADES, and other baselines. To highlight a few results, our best model improves normal and robust accuracy by 6% and 1% on CIFAR-100 compared to AT, respectively. On Imagenette, a ten-class subset of ImageNet, our model outperforms AT by 23% and 3% on normal and robust accuracy respectively.

Chawin Sitawarin, David Wagner

We study the robustness against adversarial examples of kNN classifiers and classifiers that combine kNN with neural networks. The main difficulty lies in the fact that finding an optimal attack on kNN is intractable for typical datasets. In this work, we propose a gradient-based attack on kNN and kNN-based defenses, inspired by the previous work by Sitawarin & Wagner [1]. We demonstrate that our attack outperforms their method on all of the models we tested with only a minimal increase in the computation time. The attack also beats the state-of-the-art attack [2] on kNN when k > 1 using less than 1% of its running time. We hope that this attack can be used as a new baseline for evaluating the robustness of kNN and its variants.

Grant Ho, Asaf Cidon, Lior Gavish et al.

We present the first large-scale characterization of lateral phishing attacks, based on a dataset of 113 million employee-sent emails from 92 enterprise organizations. In a lateral phishing attack, adversaries leverage a compromised enterprise account to send phishing emails to other users, benefitting from both the implicit trust and the information in the hijacked user's account. We develop a classifier that finds hundreds of real-world lateral phishing emails, while generating under four false positives per every one-million employee-sent emails. Drawing on the attacks we detect, as well as a corpus of user-reported incidents, we quantify the scale of lateral phishing, identify several thematic content and recipient targeting strategies that attackers follow, illuminate two types of sophisticated behaviors that attackers exhibit, and estimate the success rate of these attacks. Collectively, these results expand our mental models of the 'enterprise attacker' and shed light on the current state of enterprise phishing attacks.

Steven Chen, Nicholas Carlini, David Wagner

The problem of adversarial examples, evasion attacks on machine learning classifiers, has proven extremely difficult to solve. This is true even when, as is the case in many practical settings, the classifier is hosted as a remote service and so the adversary does not have direct access to the model parameters. This paper argues that in such settings, defenders have a much larger space of actions than have been previously explored. Specifically, we deviate from the implicit assumption made by prior work that a defense must be a stateless function that operates on individual examples, and explore the possibility for stateful defenses. To begin, we develop a defense designed to detect the process of adversarial example generation. By keeping a history of the past queries, a defender can try to identify when a sequence of queries appears to be for the purpose of generating an adversarial example. We then introduce query blinding, a new class of attacks designed to bypass defenses that rely on such a defense approach. We believe that expanding the study of adversarial examples from stateless classifiers to stateful systems is not only more realistic for many black-box settings, but also gives the defender a much-needed advantage in responding to the adversary.

Chawin Sitawarin, David Wagner

Robustness is an increasingly important property of machine learning models as they become more and more prevalent. We propose a defense against adversarial examples based on a k-nearest neighbor (kNN) on the intermediate activation of neural networks. Our scheme surpasses state-of-the-art defenses on MNIST and CIFAR-10 against l2-perturbation by a significant margin. With our models, the mean perturbation norm required to fool our MNIST model is 3.07 and 2.30 on CIFAR-10. Additionally, we propose a simple certifiable lower bound on the l2-norm of the adversarial perturbation using a more specific version of our scheme, a 1-NN on representations learned by a Lipschitz network. Our model provides a nontrivial average lower bound of the perturbation norm, comparable to other schemes on MNIST with similar clean accuracy.

Chawin Sitawarin, David Wagner

Despite a large amount of attention on adversarial examples, very few works have demonstrated an effective defense against this threat. We examine Deep k-Nearest Neighbor (DkNN), a proposed defense that combines k-Nearest Neighbor (kNN) and deep learning to improve the model's robustness to adversarial examples. It is challenging to evaluate the robustness of this scheme due to a lack of efficient algorithm for attacking kNN classifiers with large k and high-dimensional data. We propose a heuristic attack that allows us to use gradient descent to find adversarial examples for kNN classifiers, and then apply it to attack the DkNN defense as well. Results suggest that our attack is moderately stronger than any naive attack on kNN and significantly outperforms other attacks on DkNN.

Anish Athalye, Nicholas Carlini, David Wagner

We identify obfuscated gradients, a kind of gradient masking, as a phenomenon that leads to a false sense of security in defenses against adversarial examples. While defenses that cause obfuscated gradients appear to defeat iterative optimization-based attacks, we find defenses relying on this effect can be circumvented. We describe characteristic behaviors of defenses exhibiting the effect, and for each of the three types of obfuscated gradients we discover, we develop attack techniques to overcome it. In a case study, examining non-certified white-box-secure defenses at ICLR 2018, we find obfuscated gradients are a common occurrence, with 7 of 9 defenses relying on obfuscated gradients. Our new attacks successfully circumvent 6 completely, and 1 partially, in the original threat model each paper considers.

Nicholas Carlini, David Wagner

We construct targeted audio adversarial examples on automatic speech recognition. Given any audio waveform, we can produce another that is over 99.9% similar, but transcribes as any phrase we choose (recognizing up to 50 characters per second of audio). We apply our white-box iterative optimization-based attack to Mozilla's implementation DeepSpeech end-to-end, and show it has a 100% success rate. The feasibility of this attack introduce a new domain to study adversarial examples.

Nicholas Carlini, David Wagner

MagNet and "Efficient Defenses..." were recently proposed as a defense to adversarial examples. We find that we can construct adversarial examples that defeat these defenses with only a slight increase in distortion.

Nicholas Carlini, David Wagner

Neural networks are known to be vulnerable to adversarial examples: inputs that are close to natural inputs but classified incorrectly. In order to better understand the space of adversarial examples, we survey ten recent proposals that are designed for detection and compare their efficacy. We show that all can be defeated by constructing new loss functions. We conclude that adversarial examples are significantly harder to detect than previously appreciated, and the properties believed to be intrinsic to adversarial examples are in fact not. Finally, we propose several simple guidelines for evaluating future proposed defenses.

Primal Wijesekera, Arjun Baokar, Lynn Tsai et al.

Current smartphone operating systems regulate application permissions by prompting users on an ask-on-first-use basis. Prior research has shown that this method is ineffective because it fails to account for context: the circumstances under which an application first requests access to data may be vastly different than the circumstances under which it subsequently requests access. We performed a longitudinal 131-person field study to analyze the contextuality behind user privacy decisions to regulate access to sensitive resources. We built a classifier to make privacy decisions on the user's behalf by detecting when context has changed and, when necessary, inferring privacy preferences based on the user's past decisions and behavior. Our goal is to automatically grant appropriate resource requests without further user intervention, deny inappropriate requests, and only prompt the user when the system is uncertain of the user's preferences. We show that our approach can accurately predict users' privacy decisions 96.8% of the time, which is a four-fold reduction in error rate compared to current systems.

Nicholas Carlini, David Wagner

Neural networks provide state-of-the-art results for most machine learning tasks. Unfortunately, neural networks are vulnerable to adversarial examples: given an input $x$ and any target classification $t$, it is possible to find a new input $x'$ that is similar to $x$ but classified as $t$. This makes it difficult to apply neural networks in security-critical areas. Defensive distillation is a recently proposed approach that can take an arbitrary neural network, and increase its robustness, reducing the success rate of current attacks' ability to find adversarial examples from $95\%$ to $0.5\%$. In this paper, we demonstrate that defensive distillation does not significantly increase the robustness of neural networks by introducing three new attack algorithms that are successful on both distilled and undistilled neural networks with $100\%$ probability. Our attacks are tailored to three distance metrics used previously in the literature, and when compared to previous adversarial example generation algorithms, our attacks are often much more effective (and never worse). Furthermore, we propose using high-confidence adversarial examples in a simple transferability test we show can also be used to break defensive distillation. We hope our attacks will be used as a benchmark in future defense attempts to create neural networks that resist adversarial examples.

Michael McCoyd, David Wagner

Machine learning is increasingly used to make sense of the physical world yet may suffer from adversarial manipulation. We examine the Viola-Jones 2D face detection algorithm to study whether images can be created that humans do not notice as faces yet the algorithm detects as faces. We show that it is possible to construct images that Viola-Jones recognizes as containing faces yet no human would consider a face. Moreover, we show that it is possible to construct images that fool facial detection even when they are printed and then photographed.

Nicholas Carlini, David Wagner

We show that defensive distillation is not secure: it is no more resistant to targeted misclassification attacks than unprotected neural networks.

Linda Lee, Serge Egelman, Joong Hwa Lee et al.

Wearable devices, or "wearables," bring great benefits but also potential risks that could expose users' activities with- out their awareness or consent. In this paper, we report findings from the first large-scale survey conducted to investigate user security and privacy concerns regarding wearables. We surveyed 1,782 Internet users in order to identify risks that are particularly concerning to them; these risks are inspired by the sensor inputs and applications of popular wearable technologies. During this experiment, our questions controlled for the effects of what data was being accessed and with whom it was being shared. We also investigated how these emergent threats compared to existent mobile threats, how upcoming capabilities and artifacts compared to existing technologies, and how users ranked technical and nontechnical concerns to sketch a concrete and broad view of the wearable device landscape. We hope that this work will inform the design of future user notification, permission management, and access control schemes for wearables.

Primal Wijesekera, Arjun Baokar, Ashkan Hosseini et al.

Due to the amount of data that smartphone applications can potentially access, platforms enforce permission systems that allow users to regulate how applications access protected resources. If users are asked to make security decisions too frequently and in benign situations, they may become habituated and approve all future requests without regard for the consequences. If they are asked to make too few security decisions, they may become concerned that the platform is revealing too much sensitive information. To explore this tradeoff, we instrumented the Android platform to collect data regarding how often and under what circumstances smartphone applications are accessing protected resources regulated by permissions. We performed a 36-person field study to explore the notion of "contextual integrity," that is, how often are applications accessing protected resources when users are not expecting it? Based on our collection of 27 million data points and exit interviews with participants, we examine the situations in which users would like the ability to deny applications access to protected resources. We found out that at least 80% of our participants would have preferred to prevent at least one permission request, and overall, they thought that over a third of requests were invasive and desired a mechanism to block them.