Hyrum S. Anderson

Nicholas Carlini, Matthew Jagielski, Christopher A. Choquette-Choo et al. · deepmind, eth-zurich

Deep learning models are often trained on distributed, web-scale datasets crawled from the internet. In this paper, we introduce two new dataset poisoning attacks that intentionally introduce malicious examples to a model's performance. Our attacks are immediately practical and could, today, poison 10 popular datasets. Our first attack, split-view poisoning, exploits the mutable nature of internet content to ensure a dataset annotator's initial view of the dataset differs from the view downloaded by subsequent clients. By exploiting specific invalid trust assumptions, we show how we could have poisoned 0.01% of the LAION-400M or COYO-700M datasets for just $60 USD. Our second attack, frontrunning poisoning, targets web-scale datasets that periodically snapshot crowd-sourced content -- such as Wikipedia -- where an attacker only needs a time-limited window to inject malicious examples. In light of both attacks, we notify the maintainers of each affected dataset and recommended several low-overhead defenses.

Giovanni Apruzzese, Hyrum S. Anderson, Savino Dambra et al.

Recent years have seen a proliferation of research on adversarial machine learning. Numerous papers demonstrate powerful algorithmic attacks against a wide variety of machine learning (ML) models, and numerous other papers propose defenses that can withstand most attacks. However, abundant real-world evidence suggests that actual attackers use simple tactics to subvert ML-driven systems, and as a result security practitioners have not prioritized adversarial ML defenses. Motivated by the apparent gap between researchers and practitioners, this position paper aims to bridge the two domains. We first present three real-world case studies from which we can glean practical insights unknown or neglected in research. Next we analyze all adversarial ML papers recently published in top security conferences, highlighting positive trends and blind spots. Finally, we state positions on precise and cost-driven threat modeling, collaboration between industry and academia, and reproducible research. We believe that our positions, if adopted, will increase the real-world impact of future endeavours in adversarial ML, bringing both researchers and practitioners closer to their shared goal of improving the security of ML systems.

Elizabeth Merkhofer, Deepesh Chaudhari, Hyrum S. Anderson et al.

We present the findings of the Machine Learning Model Attribution Challenge. Fine-tuned machine learning models may derive from other trained models without obvious attribution characteristics. In this challenge, participants identify the publicly-available base models that underlie a set of anonymous, fine-tuned large language models (LLMs) using only textual output of the models. Contestants aim to correctly attribute the most fine-tuned models, with ties broken in the favor of contestants whose solutions use fewer calls to the fine-tuned models' API. The most successful approaches were manual, as participants observed similarities between model outputs and developed attribution heuristics based on public documentation of the base models, though several teams also submitted automated, statistical solutions.

Anay Mehrotra, Manolis Zampetakis, Paul Kassianik et al.

While Large Language Models (LLMs) display versatile functionality, they continue to generate harmful, biased, and toxic content, as demonstrated by the prevalence of human-designed jailbreaks. In this work, we present Tree of Attacks with Pruning (TAP), an automated method for generating jailbreaks that only requires black-box access to the target LLM. TAP utilizes an attacker LLM to iteratively refine candidate (attack) prompts until one of the refined prompts jailbreaks the target. In addition, before sending prompts to the target, TAP assesses them and prunes the ones unlikely to result in jailbreaks, reducing the number of queries sent to the target LLM. In empirical evaluations, we observe that TAP generates prompts that jailbreak state-of-the-art LLMs (including GPT4-Turbo and GPT4o) for more than 80% of the prompts. This significantly improves upon the previous state-of-the-art black-box methods for generating jailbreaks while using a smaller number of queries than them. Furthermore, TAP is also capable of jailbreaking LLMs protected by state-of-the-art guardrails, e.g., LlamaGuard.

Edward Raff, William Fleshman, Richard Zak et al.

Recent works within machine learning have been tackling inputs of ever-increasing size, with cybersecurity presenting sequence classification problems of particularly extreme lengths. In the case of Windows executable malware detection, inputs may exceed $100$ MB, which corresponds to a time series with $T=100,000,000$ steps. To date, the closest approach to handling such a task is MalConv, a convolutional neural network capable of processing up to $T=2,000,000$ steps. The $\mathcal{O}(T)$ memory of CNNs has prevented further application of CNNs to malware. In this work, we develop a new approach to temporal max pooling that makes the required memory invariant to the sequence length $T$. This makes MalConv $116\times$ more memory efficient, and up to $25.8\times$ faster to train on its original dataset, while removing the input length restrictions to MalConv. We re-invest these gains into improving the MalConv architecture by developing a new Global Channel Gating design, giving us an attention mechanism capable of learning feature interactions across 100 million time steps in an efficient manner, a capability lacked by the original MalConv CNN. Our implementation can be found at https://github.com/NeuromorphicComputationResearchProgram/MalConv2

Hyrum S. Anderson, Phil Roth

This paper describes EMBER: a labeled benchmark dataset for training machine learning models to statically detect malicious Windows portable executable files. The dataset includes features extracted from 1.1M binary files: 900K training samples (300K malicious, 300K benign, 300K unlabeled) and 200K test samples (100K malicious, 100K benign). To accompany the dataset, we also release open source code for extracting features from additional binaries so that additional sample features can be appended to the dataset. This dataset fills a void in the information security machine learning community: a benign/malicious dataset that is large, open and general enough to cover several interesting use cases. We enumerate several use cases that we considered when structuring the dataset. Additionally, we demonstrate one use case wherein we compare a baseline gradient boosted decision tree model trained using LightGBM with default settings to MalConv, a recently published end-to-end (featureless) deep learning model for malware detection. Results show that even without hyper-parameter optimization, the baseline EMBER model outperforms MalConv. The authors hope that the dataset, code and baseline model provided by EMBER will help invigorate machine learning research for malware detection, in much the same way that benchmark datasets have advanced computer vision research.

Paul Kassianik, Baturay Saglam, Alexander Chen et al.

As transformer-based large language models (LLMs) increasingly permeate society, they have revolutionized domains such as software engineering, creative writing, and digital arts. However, their adoption in cybersecurity remains limited due to challenges like scarcity of specialized training data and complexity of representing cybersecurity-specific knowledge. To address these gaps, we present Foundation-Sec-8B, a cybersecurity-focused LLM built on the Llama 3.1 architecture and enhanced through continued pretraining on a carefully curated cybersecurity corpus. We evaluate Foundation-Sec-8B across both established and new cybersecurity benchmarks, showing that it matches Llama 3.1-70B and GPT-4o-mini in certain cybersecurity-specific tasks. By releasing our model to the public, we aim to accelerate progress and adoption of AI-driven tools in both public and private cybersecurity contexts.

Robert J. Joyce, Gideon Miller, Phil Roth et al.

A lack of accessible data has historically restricted malware analysis research, and practitioners have relied heavily on datasets provided by industry sources to advance. Existing public datasets are limited by narrow scope - most include files targeting a single platform, have labels supporting just one type of malware classification task, and make no effort to capture the evasive files that make malware detection difficult in practice. We present EMBER2024, a new dataset that enables holistic evaluation of malware classifiers. Created in collaboration with the authors of EMBER2017 and EMBER2018, the EMBER2024 dataset includes hashes, metadata, feature vectors, and labels for more than 3.2 million files from six file formats. Our dataset supports the training and evaluation of machine learning models on seven malware classification tasks, including malware detection, malware family classification, and malware behavior identification. EMBER2024 is the first to include a collection of malicious files that initially went undetected by a set of antivirus products, creating a "challenge" set to assess classifier performance against evasive malware. This work also introduces EMBER feature version 3, with added support for several new feature types. We are releasing the EMBER2024 dataset to promote reproducibility and empower researchers in the pursuit of new malware research topics.

Saskia Laura Schröer, Giovanni Apruzzese, Soheil Human et al.

Our society increasingly benefits from Artificial Intelligence (AI). Unfortunately, more and more evidence shows that AI is also used for offensive purposes. Prior works have revealed various examples of use cases in which the deployment of AI can lead to violation of security and privacy objectives. No extant work, however, has been able to draw a holistic picture of the offensive potential of AI. In this SoK paper we seek to lay the ground for a systematic analysis of the heterogeneous capabilities of offensive AI. In particular we (i) account for AI risks to both humans and systems while (ii) consolidating and distilling knowledge from academic literature, expert opinions, industrial venues, as well as laypeople -- all of which being valuable sources of information on offensive AI. To enable alignment of such diverse sources of knowledge, we devise a common set of criteria reflecting essential technological factors related to offensive AI. With the help of such criteria, we systematically analyze: 95 research papers; 38 InfoSec briefings (from, e.g., BlackHat); the responses of a user study (N=549) entailing individuals with diverse backgrounds and expertise; and the opinion of 12 experts. Our contributions not only reveal concerning ways (some of which overlooked by prior work) in which AI can be offensively used today, but also represent a foothold to address this threat in the years to come.

Mayana Pereira, Rahul Dodhia, Hyrum Anderson et al.

Child Sexual Abuse Media (CSAM) is any visual record of a sexually-explicit activity involving minors. CSAM impacts victims differently from the actual abuse because the distribution never ends, and images are permanent. Machine learning-based solutions can help law enforcement quickly identify CSAM and block digital distribution. However, collecting CSAM imagery to train machine learning models has many ethical and legal constraints, creating a barrier to research development. With such restrictions in place, the development of CSAM machine learning detection systems based on file metadata uncovers several opportunities. Metadata is not a record of a crime, and it does not have legal restrictions. Therefore, investing in detection systems based on metadata can increase the rate of discovery of CSAM and help thousands of victims. We propose a framework for training and evaluating deployment-ready machine learning models for CSAM identification. Our framework provides guidelines to evaluate CSAM detection models against intelligent adversaries and models' performance with open data. We apply the proposed framework to the problem of CSAM detection based on file paths. In our experiments, the best-performing model is based on convolutional neural networks and achieves an accuracy of 0.97. Our evaluation shows that the CNN model is robust against offenders actively trying to evade detection by evaluating the model against adversarially modified data. Experiments with open datasets confirm that the model generalizes well and is deployment-ready.

Edward Raff, William Fleming, Richard Zak et al.

N-grams have been a common tool for information retrieval and machine learning applications for decades. In nearly all previous works, only a few values of $n$ are tested, with $n > 6$ being exceedingly rare. Larger values of $n$ are not tested due to computational burden or the fear of overfitting. In this work, we present a method to find the top-$k$ most frequent $n$-grams that is 60$\times$ faster for small $n$, and can tackle large $n\geq1024$. Despite the unprecedented size of $n$ considered, we show how these features still have predictive ability for malware classification tasks. More important, large $n$-grams provide benefits in producing features that are interpretable by malware analysis, and can be used to create general purpose signatures compatible with industry standard tools like Yara. Furthermore, the counts of common $n$-grams in a file may be added as features to publicly available human-engineered features that rival efficacy of professionally-developed features when used to train gradient-boosted decision tree models on the EMBER dataset.

Jonathan Woodbridge, Hyrum S. Anderson, Anjum Ahuja et al.

A homoglyph (name spoofing) attack is a common technique used by adversaries to obfuscate file and domain names. This technique creates process or domain names that are visually similar to legitimate and recognized names. For instance, an attacker may create malware with the name svch0st.exe so that in a visual inspection of running processes or a directory listing, the process or file name might be mistaken as the Windows system process svchost.exe. There has been limited published research on detecting homoglyph attacks. Current approaches rely on string comparison algorithms (such as Levenshtein distance) that result in computationally heavy solutions with a high number of false positives. In addition, there is a deficiency in the number of publicly available datasets for reproducible research, with most datasets focused on phishing attacks, in which homoglyphs are not always used. This paper presents a fundamentally different solution to this problem using a Siamese convolutional neural network (CNN). Rather than leveraging similarity based on character swaps and deletions, this technique uses a learned metric on strings rendered as images: a CNN learns features that are optimized to detect visual similarity of the rendered strings. The trained model is used to convert thousands of potentially targeted process or domain names to feature vectors. These feature vectors are indexed using randomized KD-Trees to make similarity searches extremely fast with minimal computational processing. This technique shows a considerable 13% to 45% improvement over baseline techniques in terms of area under the receiver operating characteristic curve (ROC AUC). In addition, we provide both code and data to further future research.

Miles Brundage, Shahar Avin, Jack Clark et al.

This report surveys the landscape of potential security threats from malicious uses of AI, and proposes ways to better forecast, prevent, and mitigate these threats. After analyzing the ways in which AI may influence the threat landscape in the digital, physical, and political domains, we make four high-level recommendations for AI researchers and other stakeholders. We also suggest several promising areas for further research that could expand the portfolio of defenses, or make attacks less effective or harder to execute. Finally, we discuss, but do not conclusively resolve, the long-term equilibrium of attackers and defenders.

Hyrum S. Anderson, Anant Kharkar, Bobby Filar et al.

Machine learning is a popular approach to signatureless malware detection because it can generalize to never-before-seen malware families and polymorphic strains. This has resulted in its practical use for either primary detection engines or for supplementary heuristic detection by anti-malware vendors. Recent work in adversarial machine learning has shown that deep learning models are susceptible to gradient-based attacks, whereas non-differentiable models that report a score can be attacked by genetic algorithms that aim to systematically reduce the score. We propose a more general framework based on reinforcement learning (RL) for attacking static portable executable (PE) anti-malware engines. The general framework does not require a differentiable model nor does it require the engine to produce a score. Instead, an RL agent is equipped with a set of functionality-preserving operations that it may perform on the PE file. Through a series of games played against the anti-malware engine, it learns which sequences of operations are likely to result in evading the detector for any given malware sample. This enables completely black-box attacks against static PE anti-malware, and produces functional evasive malware samples as a direct result. We show in experiments that our method can attack a gradient-boosted machine learning model with evasion rates that are substantial and appear to be strongly dependent on the dataset. We demonstrate that attacks against this model appear to also evade components of publicly hosted antivirus engines. Adversarial training results are also presented: by retraining the model on evasive ransomware samples, a subsequent attack is 33% less effective. However, there are overfitting dangers when adversarial training, which we note. We release code to allow researchers to reproduce and improve this approach.

Jonathan Woodbridge, Hyrum S. Anderson, Anjum Ahuja et al.

Various families of malware use domain generation algorithms (DGAs) to generate a large number of pseudo-random domain names to connect to a command and control (C&C) server. In order to block DGA C&C traffic, security organizations must first discover the algorithm by reverse engineering malware samples, then generating a list of domains for a given seed. The domains are then either preregistered or published in a DNS blacklist. This process is not only tedious, but can be readily circumvented by malware authors using a large number of seeds in algorithms with multivariate recurrence properties (e.g., banjori) or by using a dynamic list of seeds (e.g., bedep). Another technique to stop malware from using DGAs is to intercept DNS queries on a network and predict whether domains are DGA generated. Such a technique will alert network administrators to the presence of malware on their networks. In addition, if the predictor can also accurately predict the family of DGAs, then network administrators can also be alerted to the type of malware that is on their networks. This paper presents a DGA classifier that leverages long short-term memory (LSTM) networks to predict DGAs and their respective families without the need for a priori feature extraction. Results are significantly better than state-of-the-art techniques, providing 0.9993 area under the receiver operating characteristic curve for binary classification and a micro-averaged F1 score of 0.9906. In other terms, the LSTM technique can provide a 90% detection rate with a 1:10000 false positive (FP) rate---a twenty times FP improvement over comparable methods. Experiments in this paper are run on open datasets and code snippets are provided to reproduce the results.

Hyrum S. Anderson, Jonathan Woodbridge, Bobby Filar

Many malware families utilize domain generation algorithms (DGAs) to establish command and control (C&C) connections. While there are many methods to pseudorandomly generate domains, we focus in this paper on detecting (and generating) domains on a per-domain basis which provides a simple and flexible means to detect known DGA families. Recent machine learning approaches to DGA detection have been successful on fairly simplistic DGAs, many of which produce names of fixed length. However, models trained on limited datasets are somewhat blind to new DGA variants. In this paper, we leverage the concept of generative adversarial networks to construct a deep learning based DGA that is designed to intentionally bypass a deep learning based detector. In a series of adversarial rounds, the generator learns to generate domain names that are increasingly more difficult to detect. In turn, a detector model updates its parameters to compensate for the adversarially generated domains. We test the hypothesis of whether adversarially generated domains may be used to augment training sets in order to harden other machine learning models against yet-to-be-observed DGAs. We detail solutions to several challenges in training this character-based generative adversarial network (GAN). In particular, our deep learning architecture begins as a domain name auto-encoder (encoder + decoder) trained on domains in the Alexa one million. Then the encoder and decoder are reassembled competitively in a generative adversarial network (detector + generator), with novel neural architectures and training strategies to improve convergence.