Persistent Homology Reveals Topological Compression in LLMs Under Adversarial Attacks

A recent investigation utilizes persistent homology (PH) to explore how adversarial inputs alter the geometry and topology of internal representation spaces within large language models (LLMs). The study examined six models with parameters ranging from 3.8B to 70B, focusing on two specific attacks: indirect prompt injection and backdoor fine-tuning. Researchers identified a recurring topological signature indicating that adversarial inputs lead to topological compression, simplifying the latent space by merging diverse, compact, small-scale features into fewer, more prominent, large-scale ones. This signature is independent of the model architecture. The research fills a gap in interpretability methods, which often miss the complex, high-dimensional relationships in model representations, revealing that adversarial effects possess a detectable topological shape.