Trustworthy Computer Vision Integrity Assurance for Data, Models and Inference Outputs in Multi-Contributor Pipelines

Source summary

Develop an offline, model-agnostic vision-assurance framework assessing contributed training data, model behaviour and inference records. Identify suspicious samples/contributor patterns, evaluate model integrity within declared access limits and cryptographically bind outputs to inputs, weights and processing configuration. Distinguish distribution shift from manipulation where evidence permits, reporting explanations, confidence, actions and tamper-evident history. Support COCO/YOLO and organiser ONNX/PyTorch formats; baseline checks must not require retraining, and white-box methods must degrade honestly without access. Demonstrate reproducible public/team-generated tests and declare coverage gaps.