arXiv subjects

Philippe Loubet Moundi

Publications and source records attributed to Philippe Loubet Moundi.

3 recordsLinked to original sources

arXiv2026

Influence Score and Transformers interpretability: Measure of the Effective Impact of Attention Heads at inference time

We propose an influence score to quantify the contribution of attention heads to classification decisions in Transformer-based models designed for prompt injection detection. The score combines directional influence on the logits with structural contribution within the residual stream, enabling a multi-scale analysis at the head, layer, and network levels. Applied to a DeBERTa model specialized for prompt injection detection, our framework reveals distinct decision behaviours between correct and erroneous predictions. Our method provides an effective compromise between fine-grained circuit analysis and global output-based methods, and offers a systematic way to study decision mechanisms in Transformer classifiers.

cs.CL↗

arXiv2026

Backdoor Unlearning Generalization: A Path Toward the Removal of Unknown Triggers in LLMs

Backdoor attacks in Large Language Models (LLMs) are a growing security concern, where models can generate adversary-chosen content. Existing defenses target backdoors one at a time and typically require knowledge of the trigger, leaving the defender at a structural disadvantage when unknown backdoors may exist in a model. We show that backdoor neutralization through unlearning generalizes across backdoors: training a model to ignore a single trigger can also suppress other backdoors that were never explicitly targeted. We study this phenomenon across three model families, whose backdoors were injected via pretraining or continual pretraining, by analyzing the models obtained after removing one backdoor at a time. To understand why unlearning certain backdoors induces the suppression of others, we introduce the Cross Activation Shift Distance, to quantify the distance between model changes induced by different trainings. Our results open a new direction for LLM safety as defenders could deliberately inject controlled backdoors and then remove them, leveraging cross-backdoor transfer to also suppress unknown backdoors that an attacker may have previously introduced in the model.

cs.CL↗

arXiv2021

SideLine: How Delay-Lines (May) Leak Secrets from your SoC

To meet the ever-growing need for performance in silicon devices, SoC providers have been increasingly relying on software-hardware cooperation. By controlling hardware resources such as power or clock management from the software, developers earn the possibility to build more flexible and power efficient applications. Despite the benefits, these hardware components are now exposed to software code and can potentially be misused as open-doors to jeopardize trusted environments, perform privilege escalation or steal cryptographic secrets. In this work, we introduce SideLine, a novel side-channel vector based on delay-line components widely implemented in high-end SoCs. After providing a detailed method on how to access and convert delay-line data into power consumption information, we demonstrate that these entities can be used to perform remote power side-channel attacks. We report experiments carried out on two SoCs from distinct vendors and we recount several core-vs-core attack scenarios in which an adversary process located in one processor core aims at eavesdropping the activity of a victim process located in another core. For each scenario, we demonstrate the adversary ability to fully recover the secret key of an OpenSSL AES running in the victim core. Even more detrimental, we show that these attacks are still practicable if the victim or the attacker program runs over an operating system.

cs.CR↗