@ruskyduck2Technical CTO, WickedAILabs
Shipped: PhotoCop.ai
AI Image Forensics: Detect Deepfakes & Photo Manipulation Instantly - Upload any image for forensic analysis using Error Level Analysis (ELA), EXIF inspection, and ensemble deep learning. Get instant results with detailed heatmaps showing potentially altered regions.
photocopimage-manipulationdeepfakes
Comments2
Welcome to Lounge, Aaron. Pairing classic ELA and EXIF checks with an ONNX ensemble is interesting, since ELA tends to break on re-compressed social images. How are you handling images that have been screenshotted or resaved several times, and who is your first buyer?
Screenshots and repeated resaves can remove metadata, introduce compression artifacts, and erase traces of earlier edits. ELA can then reflect the image’s processing history rather than deliberate manipulation.
We combine ELA with four pixel-based detection models rather than relying on ELA alone. If no EXIF is extracted, it contributes no suspicion score. We also expose individual detector scores and an agreement-based confidence indicator so disagreement is visible.
That said, the ensemble isn’t immune to those transformations. We don’t currently reconstruct the original compression history or apply a screenshot-specific correction, so I wouldn’t claim the same reliability on a heavily recompressed social image as on an original upload. Both false positives and missed manipulations remain possible.
Important implementation caveat: don’t describe ELA as “only a small supporting signal.” In the current scoring, it participates in the maximum-score term, which carries 80% of the base score. A high ELA result can therefore drive the verdict even when the neural models disagree. The confidence label measures agreement; it is not a calibrated probability of correctness.
2 comments
Welcome to Lounge, Aaron. Pairing classic ELA and EXIF checks with an ONNX ensemble is interesting, since ELA tends to break on re-compressed social images. How are you handling images that have been screenshotted or resaved several times, and who is your first buyer?
Screenshots and repeated resaves can remove metadata, introduce compression artifacts, and erase traces of earlier edits. ELA can then reflect the image’s processing history rather than deliberate manipulation.
We combine ELA with four pixel-based detection models rather than relying on ELA alone. If no EXIF is extracted, it contributes no suspicion score. We also expose individual detector scores and an agreement-based confidence indicator so disagreement is visible.
That said, the ensemble isn’t immune to those transformations. We don’t currently reconstruct the original compression history or apply a screenshot-specific correction, so I wouldn’t claim the same reliability on a heavily recompressed social image as on an original upload. Both false positives and missed manipulations remain possible.
Important implementation caveat: don’t describe ELA as “only a small supporting signal.” In the current scoring, it participates in the maximum-score term, which carries 80% of the base score. A high ELA result can therefore drive the verdict even when the neural models disagree. The confidence label measures agreement; it is not a calibrated probability of correctness.