Refusal refers to the functional behavior enabling safety-aligned language models to reject harmful or unethical prompts. Following the growing scientific interest in mechanistic interpretability, recent work encoded refusal behavior as a single direction in the model’s latent space; e.g., computed as the difference between the centroids of harmful and harmless prompt representations. However, emerging evidence suggests that concepts in LLMs often appear to be encoded as a low-dimensional manifold embedded in the high-dimensional latent space. Motivated by these findings, we propose a novel method leveraging Self-Organizing Maps (SOMs) to extract multiple refusal directions. To this end, we first prove that SOMs generalize the prior work's difference-in-means technique. We then train SOMs on harmful prompt representations to identify multiple neurons. By subtracting the centroid of harmless representations from each neuron, we derive a set of multiple directions expressing the refusal concept. We validate our method on an extensive experimental setup, demonstrating that ablating multiple directions from models' internals outperforms not only the single-direction baseline but also specialized jailbreak algorithms, leading to an effective suppression of refusal. Finally, we conclude by analyzing the mechanistic implications of our approach.

Abstract To address the extremely concerning problem of software vulnerability, system security is often entrusted to Machine Learning (ML) algorithms. Despite their now established detection capabilities, such models are limited by design to flagging the entire input source code function as vulnerable, rather than precisely localizing the concerned code lines. However, the detection granularity is crucial to support human operators during software development, ensuring that such predictions reflect the true code semantics to help debug, evaluate, and fix the detected vulnerabilities. To address this issue, recent work made progress toward improving the detector’s localization ability, thus narrowing down the vulnerability detection “window” and providing more fine-grained predictions. Such approaches, however, implicitly disregard the presence of spurious correlations and biases in the data, which often predominantly influence the performance of ML algorithms. In this work, we investigate how detectors comply with this requirement by proposing an explainability-based evaluation procedure. Our approach, defined as Detection Alignment (DA), quantifies the agreement between the input source code lines that most influence the prediction and the actual localization of the vulnerability as per the ground truth. Through DA, which is model-agnostic and adaptable to different detection tasks, not limited to our use case, we analyze multiple learning-based vulnerability detectors and datasets. As a result, we show how the predictions of such models are consistently biased by non-vulnerable lines, ultimately highlighting the high impact of biases and spurious correlations.

In learning problems, the noise inherent to the task at hand hinders the possibility to infer without a certain degree of uncertainty. Quantifying this uncertainty, regardless of its wide use, assumes high relevance for security-sensitive applications. Within these scenarios, it becomes fundamental to guarantee good (i.e., trustworthy) uncertainty measures, which downstream modules can securely employ to drive the final decision-making process. However, an attacker may be interested in forcing the system to produce either (i) highly uncertain outputs jeopardizing the system’s availability or (ii) low uncertainty estimates, making the system accept uncertain samples that would instead require a careful inspection (e.g., human intervention). Therefore, it becomes fundamental to understand how to obtain robust uncertainty estimates against these kinds of attacks. In this work, we reveal both empirically and theoretically that defending against adversarial examples, i.e., carefully perturbed samples that cause misclassification, additionally guarantees a more secure, trustworthy uncertainty estimate under common attack scenarios without the need for an ad-hoc defense strategy. To support our claims, we evaluate multiple adversarial-robust classification models from the publicly available benchmark RobustBench on the CIFAR-10 and ImageNet datasets, and on a robust semantic segmentation model evaluated on Pascal-VOC. The code for the reproducibility of the experiments is available at the following link: https://github.com/pralab/UncertaintyAdversarialRobustness .

Malware evolves rapidly, forcing machine learning-based detectors to be continuously updated. With antivirus vendors processing hundreds of thousands of new samples daily, datasets can grow to billions of examples, making full retraining impractical. Continual learning (CL) has emerged as a scalable alternative, enabling incremental updates without full data access while mitigating catastrophic forgetting. In this work, we analyze a critical yet overlooked issue in this context: security regression. Unlike forgetting, which manifests as a drop in average performance on previously seen data, security regression captures harmful sample-level prediction changes, e.g., malware samples that were correctly detected before an update but evade detection afterward. This poses serious risks in security-critical applications, as the silent reintroduction of previously detected threats may undermine users' trust in the update process, leading them to perceive a regression in security even if the average model performance has actually improved. We first formalize and quantify security regression in CL-based malware detectors, revealing that up to 3-6% of malware experience it after model updates. We then address this issue by introducing a regression-aware framework to the CL setting. Specifically, we instantiate it via Positive Congruent Training (PCT), showing seamless integration with any prior CL strategy. Experiments on the ELSA, Tesseract, and AZ-Class datasets show that our method effectively halves regression across different CL scenarios while maintaining strong detection performance over time.

Gradient-based attacks are a primary tool to evaluate robustness of machine-learning models. However, many attacks tend to provide overly-optimistic evaluations as they use fixed loss functions, optimizers, step-size schedulers, and default hyperparameters. In this work, we tackle these limitations by proposing a parametric variation of the well-known fast minimum-norm attack algorithm, whose loss, optimizer, step-size scheduler, and hyperparameters can be dynamically adjusted. We re-evaluate 12 robust models, showing that our attack finds smaller adversarial perturbations without requiring any additional tuning. This also enables reporting adversarial robustness as a function of the perturbation budget, providing a more complete evaluation than that offered by fixed-budget attacks, while remaining efficient. We release our open-source code at https://github.com/pralab/HO-FMN . • We improve minimum-norm attacks using different losses, optimizers, and schedulers. • We leverage hyperparameter optimization to improve the attack performance. • We compute full robustness-perturbation curves with a single, effective attack.

Recent work has proposed neural network pruning techniques to reduce the size of a network while preserving robustness against adversarial examples, i.e., well-crafted inputs inducing a misclassification. These methods, which we refer to as adversarial pruning methods, involve complex and articulated designs, making it difficult to analyze the differences and establish a fair and accurate comparison. In this work, we overcome these issues by surveying current adversarial pruning methods and proposing a novel robustness-oriented taxonomy to categorize them based on two main dimensions: the pipeline , defining when to prune; and the specifics , defining how to prune. We then highlight the limitations of current empirical analyses and propose a novel, fair evaluation benchmark to address them. We finally conduct an empirical re-evaluation of current adversarial pruning methods and discuss the results, highlighting the shared traits of top-performing adversarial pruning methods, as well as common issues. We welcome contributions in our publicly-available benchmark at https://github.com/pralab/AdversarialPruningBenchmark .

Evaluating the adversarial robustness of machine-learning models using gradient-based attacks is challenging.In this work, we show that hyperparameter optimization can improve fast minimum-norm attacks by automating the selection of the loss function, the optimizer, and the step-size scheduler, along with the corresponding hyperparameters.Our extensive evaluation involving several robust models demonstrates the improved efficacy of fast minimum-norm attacks when hyped up with hyperparameter optimization.We release our open-source code at https://github.com/pralab/HO-FMN.