Model inversion attacks typically exploit prediction outputs and confidence scores from AI models through iterative queries and gradient-based optimization to reconstruct sensitive training data.
Open the full topic