Model Inversion Attacks: Reconstructing Faces and Voices from Model Weights
Factos validados
Fredrikson et al. / Carnegie Mellon. Shows how attackers, through targeted queries to a trained recognition system, can reconstruct the face or biometric traits of people in the training set.
Explicação simples
The composite sketch from the shadows: Ask often enough 'Does this look like John Doe?' and you can puzzle the real passport photo of John back together from the probabilities.
Relevância de segurança e dica prática
Proves that AI models are not 'anonymous' black boxes but can mirror personal data.
Fonte primária / referência
- CMU Security Research