vLLM MultiModalHasher Image Hashing Vulnerability Leading to Hash Collisions

A vulnerability exists in the vLLM image hashing method within the MultiModalHasher class, specifically in versions 0.7.0 prior to 0.9.0. The issue arises because the hashing method serializes PIL.Image.Image objects using only the raw pixel data, excluding important metadata such as the image's dimensions and color mode. This oversight can cause different images to generate identical hash values, leading to hash collisions, incorrect cache hits, and potential data leakage or security risks.

Impact

The vulnerability can cause hash collisions, allowing different images to be treated as identical by the caching system. This could result in incorrect cache hits, abnormal responses, and data leakage.

Reproduction

The vulnerability can be reproduced by creating two images of different sizes but with the same pixel data. When these images are processed by the MultiModalHasher's hashing method, they will generate the same hash value, demonstrating the hash collision.

Remediation

The vulnerability has been addressed in vLLM version 0.9.0. Users should update to this version.