Why Are LLM Backdoor Defenses Fragmented? A Feature-Level Explanation with Sparse Autoencoders
The paper uses sparse autoencoders to mechanistically analyze LLM backdoors at the feature level, comparing clean vs. poisoned models on clean vs. triggered inputs and tracing malicious logit shifts to SAE features with roles like interaction, suppressed, mixed, and weight-modified features. Its key insight is that dirty-label and clean-label backdoors are encoded in qualitatively different feature patterns, which explains why prior defenses don’t transfer cleanly across attack types; the SAE-based feature clamping results look promising and interpretable, though the evidence seems bounded by the evaluated models/attack settings and by how faithfully the learned SAE features reflect the model’s true internal mechanisms.
Abstract
Backdoor attacks pose a serious threat to large language models (LLMs), but existing defenses remain fragmented, failing to pro?vide unified defense against both dirty-label and clean-label attacks. To investigate why such fragmentation arises, we present the first systematic feature-level mechanistic analysis of LLM backdoors using sparse autoencoders (SAEs). Starting from a 2 x 2 comparison of clean and poisoned models on clean and triggered inputs, we trace backdoor-induced logit shifts to high-contributing SAE features and categorize them into four roles: interac?tion, suppressed, mixed, and weight-modified features. This taxonomy reveals system?atic encoding differences: dirty-label back?doors are dominated by isolated interaction features, whereas clean-label backdoors rely more on heterogeneous mixtures of mixed and weight-modified features. These differ?ences explain why existing defenses remain fragmented across attack paradigms. We val?idate this hypothesis through inference-time feature clamping, which reduces ASR to at most 10.8% in most dirty-label settings and at most 15.4% in the majority of clean-label settings, while preserving benign-task perfor?mance. These results show that SAE-based analysis can explain defense fragmentation and guide interpretable backdoor mitigation.