# Why Are LLM Backdoor Defenses Fragmented? A Feature-Level Explanation with Sparse Autoencoders

arXiv: https://arxiv.org/abs/2608.30403 · HTML: https://mlsanity.com/paper/2608.30403/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Backdoor and data poisoning attacks, one of 12 directions in LLM safety and security.
- Builds on 27 of its 37 references that are in the map.
- First posted on arXiv 31 Aug 2026.

## Placed in

- LLM safety and security / [Backdoor and data poisoning attacks](https://mlsanity.com/safety/backdoor-and-data-poisoning-attacks/)

## Builds on (most cited first, 27 of 27)

- [Character-level Convolutional Networks for Text Classification](https://mlsanity.com/paper/1509.01626/) (2015)
- [Gemma 2: Improving Open Language Models at a Practical Size](https://mlsanity.com/paper/2408.00118/) (2024)
- [BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain](https://mlsanity.com/paper/1708.06733/) (2017)
- [Refusal in Language Models Is Mediated by a Single Direction](https://mlsanity.com/paper/2406.11717/) (2024)
- [STRIP: A Defence Against Trojan Attacks on Deep Neural Networks](https://mlsanity.com/paper/1902.06531/) (2019)
- [Detecting Backdoor Attacks on Deep Neural Networks by Activation Clustering](https://mlsanity.com/paper/1811.03728/) (2018)
- [Weight Poisoning Attacks on Pre-trained Models](https://mlsanity.com/paper/2004.06660/) (2020)
- [ONION: A Simple and Effective Defense Against Textual Backdoor Attacks](https://mlsanity.com/paper/2011.10369/) (2020)
- [Hidden Killer: Invisible Textual Backdoor Attacks with Syntactic Trigger](https://mlsanity.com/paper/2105.12400/) (2021)
- [BadNL: Backdoor Attacks against NLP Models with Semantic-preserving Improvements](https://mlsanity.com/paper/2006.01043/) (2020)
- [Latent Adversarial Training Improves Robustness to Persistent Harmful Behaviors in LLMs](https://mlsanity.com/paper/2407.15549/) (2024)
- [Backdooring Instruction-Tuned Large Language Models with Virtual Prompt Injection](https://mlsanity.com/paper/2307.16888/) (2023)
- [RAP: Robustness-Aware Perturbations for Defending against Backdoor Attacks on NLP Models](https://mlsanity.com/paper/2110.07831/) (2021)
- [On the Role of Attention Heads in Large Language Model Safety](https://mlsanity.com/paper/2410.13708/) (2024)
- [Triggerless Backdoor Attack for NLP Tasks with Clean Labels](https://mlsanity.com/paper/2111.07970/) (2021)
- [BEEAR: Embedding-based Adversarial Removal of Safety Backdoors in Instruction-tuned Language Models](https://mlsanity.com/paper/2406.17092/) (2024)
- [Towards Understanding Safety Alignment: A Mechanistic Perspective from Safety Neurons](https://mlsanity.com/paper/2406.14144/) (2024)
- [CROW: Eliminating Backdoors from Large Language Models via Internal Consistency Regularization](https://mlsanity.com/paper/2411.12768/) (2024)
- [JailbreakLens: Interpreting Jailbreak Mechanism in the Lens of Representation and Circuit](https://mlsanity.com/paper/2411.11114/) (2024)
- [BadActs: A Universal Backdoor Defense in the Activation Space](https://mlsanity.com/paper/2405.11227/) (2024)
- [BadJudge: Backdoor Vulnerabilities of LLM-as-a-Judge](https://mlsanity.com/paper/2503.00596/) (2025)
- [CorrSteer: Generation-Time LLM Steering via Correlated Sparse Autoencoder Features](https://mlsanity.com/paper/2508.12535/) (2025)
- [Backdoor Attribution: Elucidating and Controlling Backdoor in Language Models](https://mlsanity.com/paper/2509.21761/) (2025)
- [Sparse Autoencoders are Capable LLM Jailbreak Mitigators](https://mlsanity.com/paper/2602.12418/) (2026)
- [Defending Code Language Models against Backdoor Attacks with Deceptive Cross-Entropy Loss](https://mlsanity.com/paper/2407.08956/) (2024)
- [Interpretable Safety Alignment via SAE-Constructed Low-Rank Subspace Adaptation](https://mlsanity.com/paper/2512.23260/) (2025)
- [MirageBackdoor: A Stealthy Attack that Induces Think-Well-Answer-Wrong Reasoning](https://mlsanity.com/paper/2604.06840/) (2026)

## Abstract (arXiv)

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.