Fabio Brau

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.

Federated Learning has been popularized in recent years for applications involving personal or sensitive data, as it allows the collaborative training of machine learning models through local updates at the data-owners' premises, which does not require the sharing of the data itself. Considering the risk of leakage or misuse by any of the data-owners, many works attempt to protect their copyright, or even trace the origin of a potential leak through unique watermarks identifying each participant's model copy. Realistic accusation scenarios impose a black-box setting, where watermarks are typically embedded as a set of sample-label pairs. The threat of collusion, however, where multiple bad actors conspire together to produce an untraceable model, has been rarely addressed, and previous works have been limited to shallow networks and near-linearly separable main tasks. To the best of our knowledge, this work is the first to present a general collusion-resistant embedding method for black-box traitor tracing in Federated Learning: BlackCATT, which introduces a novel collusion-aware embedding loss term and, instead of using a fixed trigger set, iteratively optimizes the triggers to aid convergence and traitor tracing performance. Experimental results confirm the efficacy of the proposed scheme across different architectures and datasets. Furthermore, for models that would otherwise suffer from update incompatibility on the main task after learning different watermarks (e.g., architectures including batch normalization layers), our proposed BlackCATT+FR incorporates functional regularization through a set of auxiliary examples at the aggregator, promoting a shared feature space among model copies without compromising traitor tracing performance.

Adversarial pruning methods have emerged as a powerful tool for compressing neural networks while preserving robustness against adversarial attacks. These methods typically follow a three-step pipeline: (i) pretrain a robust model, (ii) select a binary mask for weight pruning, and (iii) finetune the pruned model. To select the binary mask, these methods minimize a robust loss by assigning an importance score to each weight, and then keep the weights with the highest scores. However, this score-space optimization can lead to sharp local minima in the robust loss landscape and, in turn, to an unstable mask selection, reducing the robustness of adversarial pruning methods. To overcome this issue, we propose a novel plug-in method for adversarial pruning, termed Score-space Sharpness-aware Adversarial Pruning (S2AP). Through our method, we introduce the concept of score-space sharpness minimization, which operates during the mask search by perturbing importance scores and minimizing the corresponding robust loss. Extensive experiments across various datasets, models, and sparsity levels demonstrate that S2AP effectively minimizes sharpness in score space, stabilizing the mask selection, and ultimately improving the robustness of adversarial pruning methods.