SPECTER: Hunting Ghosts in the KV Cache

This site does not allow itself to be embedded. You can still read it on the original site — the toolbar below keeps your place in the directory.

Two research domains have been studying KV cache attacks and quantization exploitation separately. SPECTER bridges the gap, testing whether crafted input tokens can exploit quantization boundaries in the KV cache to manipulate model behavior through normal inference.

Read on blog.digitalgrease.dev ↗