Update: V5 Western Blot Detection Model
We completed training and evaluation of V5, an updated Western blot manipulation detection model featuring a redesigned residual CNN architecture and improved training methods. On a paper-disjoint benchmark of synthetic manipulations under realistic image-processing conditions, V5 achieved 62.6% recall for AI-generated band splices, compared with 39.3% for V3. Its AUROC improved from 0.773 to 0.872, while the clean-image false-positive rate decreased from 7.0% to 4.6%. On a separate test using an AI-generated source image withheld from training, V5 detected 52.5% of manipulated images, compared with 30.8% for V3. Performance remained substantially lower under aggressive resizing, and further validation on real-world cases is needed. Training and evaluation finished after the hackathon submission deadline, so these results were not available for our original submission. V5 represents an improvement over our previous experimental model, while scientific image duplication detection remains the project's primary focus.
Log in or sign up for Devpost to join the conversation.