MorphLens - Exploiting Intermediate MLLM Representations for Single-Image Morph Attack Detection
Nitish Shukla et al.; Under review
Research Goal
Morph attack detection decides whether a single face image is a morph that could match two people. Multimodal LLMs already show useful morph-detection ability, but existing approaches read only the model’s endpoints: its final answer or its vision-encoder features. MorphLens asks where inside the decoder the morph evidence is strongest, and finds that morph and bona fide images can separate before the final language output.
How it works
- Frozen MLLM. The face image and a fixed yes/no instruction pass once through a fully frozen multimodal LLM.
- Probe every depth. A lightweight linear probe reads pooled image-token representations at each decoder layer.
- Select and calibrate. The most informative intermediate layers are chosen on validation data, and their scores are aligned with validation percentiles.
- Fuse and classify. The calibrated scores are combined into one morph score. The probes are fit on synthetic data only and evaluated unchanged on external morph databases.
Status: under review.