Activation Differences Reveal Backdoors: A Comparison of SAE Architectures

• URL: http://arxiv.org/abs/2605.07324v1
• Date: Fri, 08 May 2026 06:30:26 GMT
• Title: Activation Differences Reveal Backdoors: A Comparison of SAE Architectures
• Abstract summary: Backdoor attacks on language models pose a significant threat to AI safety.<n>We investigate two sparse autoencoder architectures for isolating backdoor-related features in fine-tuned models.<n>We find that Diff-SAE consistently and substantially outperforms Crosscoders for backdoor isolation.
• Score: 5.218766876318545
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Backdoor attacks on language models pose a significant threat to AI safety, where models behave normally on most inputs but exhibit harmful behavior when triggered by specific patterns. Detecting such backdoors through mechanistic interpretability remains an open challenge. We investigate two sparse autoencoder architectures -- Crosscoders and Differential SAEs (Diff-SAE) -- for isolating backdoor-related features in fine-tuned models. Using a controlled SQL injection backdoor triggered by year-based context ("2024" triggers vulnerable code, "2023" triggers safe code), we evaluate both approaches across LoRA and full-rank fine-tuning regimes on SmolLM2-360M. We find that Diff-SAE consistently and substantially outperforms Crosscoders for backdoor isolation. Diff-SAE achieves a Backdoor Isolation Score (BIS) of 0.40 with perfect precision (1.0) and zero false positive rate across most experimental conditions, while Crosscoders fail almost entirely with BIS below 0.02 in most cases. This performance gap holds across multiple transformer layers (14, 18, 22, 26) and both fine-tuning regimes, with full-rank fine-tuning producing particularly clean backdoor signals. Our results suggest that backdoors manifest as directional activation shifts rather than sparse feature activations, making difference-based representations fundamentally more effective for detection. These findings have important implications for AI safety monitoring and the development of interpretability tools for detecting model manipulation.

Related papers

• Lightweight and Fast Backdoor Model Detection [36.85393813658947]
  DFBScanner is a lightweight static parameter inspection framework for fast backdoor scanning.<n>We show that DFBScanner achieves a 97.17% true-positive rate, 0.95% false-positive rate, and an average detection time of only 1 ms per model.
  arXiv  Detail & Related papers  (2026-05-17T16:02:40Z)
• Backdoor Directions in Vision Transformers [56.382912038371046]
  This paper investigates how Backdoor Attacks are represented within Vision Transformers (ViTs)<n>We identify a specific trigger direction'' in the model's activations that corresponds to the internal representation of the trigger.<n>Using this direction as a diagnostic tool, we trace how backdoor features are processed across layers.
  arXiv  Detail & Related papers  (2026-03-11T14:13:48Z)
• Backdoor4Good: Benchmarking Beneficial Uses of Backdoors in LLMs [63.668008418097905]
  We show that beneficial backdoors can achieve high controllability, tamper-resistance, and stealthiness while preserving clean-task performance.<n>Our findings demonstrate new insights that backdoors need not be inherently malicious; when properly designed, they can serve as modular, interpretable, and beneficial building blocks for trustworthy AI systems.
  arXiv  Detail & Related papers  (2026-03-08T04:01:13Z)
• Assimilation Matters: Model-level Backdoor Detection in Vision-Language Pretrained Models [71.44858461725893]
  Given a model fine-tuned by an untrusted third party, determining whether the model has been injected with a backdoor is a critical and challenging problem.<n>Existing detection methods usually rely on prior knowledge of training dataset, backdoor triggers and targets.<n>We introduce Assimilation Matters in DETection (AMDET), a novel model-level detection framework that operates without any such prior knowledge.
  arXiv  Detail & Related papers  (2025-11-29T06:20:00Z)
• Backdoor Collapse: Eliminating Unknown Threats via Known Backdoor Aggregation in Language Models [75.29749026964154]
  Ourmethod reduces the average Attack Success Rate to 4.41% across multiple benchmarks.<n>Clean accuracy and utility are preserved within 0.5% of the original model.<n>The defense generalizes across different types of backdoors, confirming its robustness in practical deployment scenarios.
  arXiv  Detail & Related papers  (2025-10-11T15:47:35Z)
• Lethe: Purifying Backdoored Large Language Models with Knowledge Dilution [49.78359632298156]
  Large language models (LLMs) have seen significant advancements, achieving superior performance in various Natural Language Processing (NLP) tasks.<n> backdoor attacks, where models behave normally for standard queries but generate harmful responses or unintended output when specific triggers are activated.<n>We present LETHE, a novel method to eliminate backdoor behaviors from LLMs through knowledge dilution.
  arXiv  Detail & Related papers  (2025-08-28T17:05:18Z)
• Trigger without Trace: Towards Stealthy Backdoor Attack on Text-to-Image Diffusion Models [70.03122709795122]
  Backdoor attacks targeting text-to-image diffusion models have advanced rapidly.<n>Current backdoor samples often exhibit two key abnormalities compared to benign samples.<n>We propose Trigger without Trace (TwT) by explicitly mitigating these consistencies.
  arXiv  Detail & Related papers  (2025-03-22T10:41:46Z)
• Neutralizing Backdoors through Information Conflicts for Large Language Models [20.6331157117675]
  We present a novel method to eliminate backdoor behaviors from large language models (LLMs)<n>We leverage a lightweight dataset to train a conflict model, which is then merged with the backdoored model to neutralize malicious behaviors.<n>We can reduce the attack success rate of advanced backdoor attacks by up to 98% while maintaining over 90% clean data accuracy.
  arXiv  Detail & Related papers  (2024-11-27T12:15:22Z)
• Towards Robust Object Detection: Identifying and Removing Backdoors via Module Inconsistency Analysis [5.8634235309501435]
  We propose a backdoor defense framework tailored to object detection models. By quantifying and analyzing inconsistencies, we develop an algorithm to detect backdoors. Experiments with state-of-the-art two-stage object detectors show our method achieves a 90% improvement in backdoor removal rate.
  arXiv  Detail & Related papers  (2024-09-24T12:58:35Z)
• Diff-Cleanse: Identifying and Mitigating Backdoor Attacks in Diffusion Models [3.134071086568745]
  Diffusion models (DMs) are regarded as one of the most advanced generative models today. Recent studies suggest that DMs are vulnerable to backdoor attacks. This vulnerability poses substantial risks, including reputational damage to model owners. We introduce Diff-Cleanse, a novel two-stage backdoor defense framework specifically designed for DMs.
  arXiv  Detail & Related papers  (2024-07-31T03:54:41Z)
• BEEAR: Embedding-based Adversarial Removal of Safety Backdoors in Instruction-tuned Language Models [57.5404308854535]
  Safety backdoor attacks in large language models (LLMs) enable the stealthy triggering of unsafe behaviors while evading detection during normal interactions. We present BEEAR, a mitigation approach leveraging the insight that backdoor triggers induce relatively uniform drifts in the model's embedding space. Our bi-level optimization method identifies universal embedding perturbations that elicit unwanted behaviors and adjusts the model parameters to reinforce safe behaviors against these perturbations.
  arXiv  Detail & Related papers  (2024-06-24T19:29:47Z)
• Exposing Backdoors in Robust Machine Learning Models [0.5672132510411463]
  We show that adversarially robust models are susceptible to backdoor attacks. backdoors are reflected in the feature representation of such models. This observation is leveraged to detect backdoor-infected models via a detection technique called AEGIS.
  arXiv  Detail & Related papers  (2020-02-25T04:45:26Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.