Backdoor Detection

Microsoft researchers unveil detection method for poisoned AI models achieving 88% accuracy with zero false positives across 47 sleeper agent models.