Ethan Rathbun

Ethan Rathbun, Kaleel Mahmood, Sohaib Ahmad et al.

Recent advances in adversarial machine learning have shown that defenses considered to be robust are actually susceptible to adversarial attacks which are specifically customized to target their weaknesses. These defenses include Barrage of Random Transforms (BaRT), Friendly Adversarial Training (FAT), Trash is Treasure (TiT) and ensemble models made up of Vision Transformers (ViTs), Big Transfer models and Spiking Neural Networks (SNNs). We first conduct a transferability analysis, to demonstrate the adversarial examples generated by customized attacks on one defense, are not often misclassified by another defense. This finding leads to two important questions. First, how can the low transferability between defenses be utilized in a game theoretic framework to improve the robustness? Second, how can an adversary within this framework develop effective multi-model attacks? In this paper, we provide a game-theoretic framework for ensemble adversarial attacks and defenses. Our framework is called Game theoretic Mixed Experts (GaME). It is designed to find the Mixed-Nash strategy for both a detector based and standard defender, when facing an attacker employing compositional adversarial attacks. We further propose three new attack algorithms, specifically designed to target defenses with randomized transformations, multi-model voting schemes, and adversarial detector architectures. These attacks serve to both strengthen defenses generated by the GaME framework and verify their robustness against unforeseen attacks. Overall, our framework and analyses advance the field of adversarial machine learning by yielding new insights into compositional attack and defense formulations.

Nuo Xu, Kaleel Mahmood, Haowen Fang et al.

Spiking neural networks (SNNs) have drawn much attention for their high energy efficiency and recent advances in classification performance. However, unlike traditional deep learning, the robustness of SNNs to adversarial examples remains underexplored. This work advances the adversarial attack side of SNNs and makes three major contributions. First, we show that successful white-box attacks on SNNs strongly depend on the surrogate gradient estimation technique, even for adversarially trained models. Second, using the best single surrogate gradient estimator, we study the transferability of adversarial examples between SNNs and state-of-the-art architectures such as Vision Transformers (ViTs) and CNNs. Our analysis reveals two major gaps: no existing white-box attack leverages multiple surrogate estimators, and no single attack effectively fools both SNNs and non-SNN models simultaneously. Third, we propose the Mixed Dynamic Spiking Estimation (MDSE) attack, which dynamically combines multiple surrogate gradients to overcome these gaps. MDSE produces adversarial examples that fool both SNN and non-SNN models, achieving up to 91.4% higher effectiveness on SNN/ViT ensembles and a 3x boost on adversarially trained SNN ensembles over Auto-PGD. Experiments span three datasets (CIFAR-10, CIFAR-100, ImageNet) and nineteen classifiers, and we will release code and models upon publication.

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.

Wo Wei Lin, Ethan Rathbun, Enrico Marchesini Xiang Zhi Tan

Multi-agent reinforcement learning (MARL) in real-world use cases may need to adapt to external natural language instructions that interrupt ongoing behavior and conflict with long-horizon objectives. However, conditioning rewards on instructions introduces a fundamental failure mode as Bellman updates couple value estimates across instruction contexts, leading to inconsistent values when instructions interrupt macro-actions. We propose Macro-Action Value Correction for Instruction Compliance (MAVIC), which corrects Bellman backups at instruction boundaries by correcting the incoming instruction objective and restoring the continuation value under the current objective. Unlike reward shaping, MAVIC modifies the bootstrapping target itself, enabling consistent value estimation under stochastic instruction switching within a unified policy. We provide theoretical analysis and an actor-critic implementation, and show that MAVIC achieves high instruction compliance while preserving base task performance in increasingly complex cooperative multi-agent environments.

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.

Jieren Deng, Aaron Palmer, Rigel Mahmood et al.

Achieving resiliency against adversarial attacks is necessary prior to deploying neural network classifiers in domains where misclassification incurs substantial costs, e.g., self-driving cars or medical imaging. Recent work has demonstrated that robustness can be transferred from an adversarially trained teacher to a student model using knowledge distillation. However, current methods perform distillation using a single adversarial and vanilla teacher and consider homogeneous architectures (i.e., residual networks) that are susceptible to misclassify examples from similar adversarial subspaces. In this work, we develop a defense framework against adversarial attacks by distilling adversarial robustness using heterogeneous teachers (DARHT). In DARHT, the student model explicitly represents teacher logits in a student-teacher feature map and leverages multiple teachers that exhibit low adversarial example transferability (i.e., exhibit high performance on dissimilar adversarial examples). Experiments on classification tasks in both white-box and black-box scenarios demonstrate that DARHT achieves state-of-the-art clean and robust accuracies when compared to competing adversarial training and distillation methods in the CIFAR-10, CIFAR-100, and Tiny ImageNet datasets. Comparisons with homogeneous and heterogeneous teacher sets suggest that leveraging teachers with low adversarial example transferability increases student model robustness.

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.

Kaleel Mahmood, Caleb Manicke, Ethan Rathbun et al.

We show the security risk associated with using machine learning classifiers in United States election tabulators. The central classification task in election tabulation is deciding whether a mark does or does not appear on a bubble associated to an alternative in a contest on the ballot. Barretto et al. (E-Vote-ID 2021) reported that convolutional neural networks are a viable option in this field, as they outperform simple feature-based classifiers. Our contributions to election security can be divided into four parts. To demonstrate and analyze the hypothetical vulnerability of machine learning models on election tabulators, we first introduce four new ballot datasets. Second, we train and test a variety of different models on our new datasets. These models include support vector machines, convolutional neural networks (a basic CNN, VGG and ResNet), and vision transformers (Twins and CaiT). Third, using our new datasets and trained models, we demonstrate that traditional white box attacks are ineffective in the voting domain due to gradient masking. Our analyses further reveal that gradient masking is a product of numerical instability. We use a modified difference of logits ratio loss to overcome this issue (Croce and Hein, ICML 2020). Fourth, in the physical world, we conduct attacks with the adversarial examples generated using our new methods. In traditional adversarial machine learning, a high (50% or greater) attack success rate is ideal. However, for certain elections, even a 5% attack success rate can flip the outcome of a race. We show such an impact is possible in the physical domain. We thoroughly discuss attack realism, and the challenges and practicality associated with printing and scanning ballot adversarial examples.

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.

Kaleel Mahmood, Rigel Mahmood, Ethan Rathbun et al.

The field of adversarial machine learning has experienced a near exponential growth in the amount of papers being produced since 2018. This massive information output has yet to be properly processed and categorized. In this paper, we seek to help alleviate this problem by systematizing the recent advances in adversarial machine learning black-box attacks since 2019. Our survey summarizes and categorizes 20 recent black-box attacks. We also present a new analysis for understanding the attack success rate with respect to the adversarial model used in each paper. Overall, our paper surveys a wide body of literature to highlight recent attack developments and organizes them into four attack categories: score based attacks, decision based attacks, transfer attacks and non-traditional attacks. Further, we provide a new mathematical framework to show exactly how attack results can fairly be compared.