3 January 2024 in Computer
We propose a novel reinforcement learning-based adversarial example attack, Adversarial Reinforcement Learning Agent, designed to learn imperceptible perturbation that causes misclassification when added to the input of a deep learning classifier.
26 August 2025 in ACM Transactions on Privacy and Security
Tor users derive anonymity in part from the size of the Tor user base, but Tor struggles to attract and support more users due to performance limitations. Previous works have proposed modifications to Tor’s path selection algorithm to enhance both performance and security, but many proposals have unintended consequences due to incorporating information related to client location. We instead propose selecting paths using a global view of the network, independent of client location, and we propose doing so with a machine learning classifier to predict the performance of a given path before building a circuit. We show through a variety of simulated and live experimental settings, across different time periods, that this approach can significantly improve performance compared to Tor’s default path selection algorithm and two previously proposed approaches. In addition to evaluating the security of our approach with traditional metrics, we propose a novel anonymity metric that captures information leakage resulting from location-aware path selection, and we show that our path selection approach leaks no more information than the default path selection algorithm.
9 November 2024 in ACM Transactions on Privacy and Security
In an increasingly digital and interconnected world, online anonymity and privacy are paramount issues for Internet users. To address this, tools like The Onion Router (Tor) offer anonymous and private communication by routing traffic through multiple relays with multiple layers of encryption. However, traffic fingerprinting attacks have threatened anonymity and privacy. In response, the community has proposed additional defenses for Tor, but fingerprinting techniques that utilize deep neural networkss (DNNs) have undermined many of these defenses. The latest defenses that are both lightweight and robust against DNNs use adversarial examples, but these defenses require either the full traffic trace beforehand or a database of pre-computed adversarial examples. We propose Prism , a defense against fingerprinting attacks that utilizes adversarial examples with neither prior access to the full traffic trace nor a database. We describe a novel method of adversarial example generation as input is learned over time. Prism injects these adversarial examples into the Tor traffic stream to prevent DNNs from accurately classifying both websites and videos that a user is viewing, even if the DNN is hardened by adversarial training. We also show that the Tor network could implement Prism entirely on relays under certain conditions, extending the defense to users who may run Tor on devices without graphics processing units.
24 March 2026 in ACM Transactions on Privacy and Security
Machine learning (ML) suffers from a persistent and critical flaw: adversarial examples. Many new forms of adversarial example attacks have been invented and many narrow defenses have been proposed. Unfortunately, no defensive approach can withstand current attacks. We hypothesize that ML model robustness can be improved with approaches that delineate the data-point-sparse latent space between data-dense regions of a model’s classification space as a barrier class. We introduce one such defense, PadNet, that builds a barrier class using a combination of training samples that mix multiple classes together. It leverages this barrier class to separate decision boundaries between benign classes with regions of padding. PadNet then implements a gradient regularization strategy that penalizes gradients in the direction of the barrier class, causing the decision boundary to draw tighter around training samples increasing boundary thickness between classes. We evaluate PadNet against a sampling of the most effective state-of-the-art attacks, demonstrating that it offers significant robustness and reliability compared to current defenses. We also test it against adaptive attacks and find that PadNet remains robust against them.
19 February 2026 in International Conference on Cyber Warfare and Security
Red teaming is a common cybersecurity practice that simulates real-world adversarial cyber operations on defended systems to identify vulnerabilities. Current red team tools often have limited logging capabilities, resulting in insufficient analysis that prevents red teams from receiving real-time feedback and insights after operations. The application of machine learning for automating the analysis of red team operations is severely constrained by the scarcity of labeled, real-world log data. This research addresses this challenge by exploring the potential of using synthetic data to train attack-detection models for Cobalt Strike logs. We systematically evaluate three different training approaches for analyzing Cobalt Strike operational logs: synthetic-only, real-world-only, and a hybrid approach that combines both data types. Our methodology employs a comprehensive feature engineering pipeline that includes both programmatic log generation for creating large-scale structured data and large language model techniques for introducing variety and edge cases. We transform each log file into a high-dimensional vector that includes event types, command verbs, temporal activity patterns, and mappings to the MITRE ATT&CK knowledge base. Random Forest classification models are trained using this feature set to distinguish between successful and failed attack scenarios. By rigorously testing each training approach against a manually labeled ground-truth set of 112 authentic Cobalt Strike logs, we quantify the performance and limitations of each strategy. Our main contribution is demonstrating that a hybrid training strategy achieves 94% accuracy, greatly surpassing synthetic-only models (56%) and real-world-only models (79%). This combined approach effectively addresses both the domain gap in synthetic data and the data scarcity in small, real-world datasets. The hybrid model learns attack diversity from over 30,000 synthetic scenarios while grounding understanding in the authentic structural patterns of real logs, providing a 15 percentage-point improvement over real-data-only approaches. This research offers a practical framework for enhancing limited real-world cybersecurity datasets by strategically integrating synthetic data, enabling immediate use in Department of Defense red team operations and wider cybersecurity machine learning applications.