Activation Differences Reveal Backdoors: A Comparison of SAE Architectures
Fuente:
arXiv
Enregistré dans:
| Auteur principal: | |
|---|---|
| Format: | Preprint |
| Publié: |
2026
|
| Sujets: | |
| Accès en ligne: | |
| Tags: |
Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
|
| _version_ | 1866909025803698176 |
|---|---|
| author | Kumar, Sachin |
| author_facet | Kumar, Sachin |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2605_07324 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Activation Differences Reveal Backdoors: A Comparison of SAE Architectures Kumar, Sachin Computation and Language Artificial Intelligence Cryptography and Security Machine Learning I.2.6; K.6.5 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. |
| title | Activation Differences Reveal Backdoors: A Comparison of SAE Architectures |
| topic | Computation and Language Artificial Intelligence Cryptography and Security Machine Learning I.2.6; K.6.5 |
| url | https://arxiv.org/abs/2605.07324 |