Model inversion: What prediction responses can reveal
Research reconstructed recognisable facial images using confidence values. Preconditions and reconstruction quality are crucial to interpreting the result.
Revisión editorial:
El texto de este artículo está disponible en inglés en esta versión del sitio.
Contexto y fuentes
Fredrikson, Jha and Ristenpart studied inversion attacks exploiting confidence information in 2015. Tested scenarios included facial recognition and sensitive survey attributes. Recognisable face representations could be reconstructed under the investigated conditions. The study also examines countermeasures, including reducing the precision of exposed confidence values. [1]
Alcance y límites
A reconstruction is not automatically the original photograph. This particular evidence does not establish arbitrary reconstruction of voices. [1]
Explicación sencilla
Repeated guesses accompanied by feedback about closeness to a target may reveal more than a single simple prediction.
Recomendaciones prácticas
Our recommendation: review the additional information exposed by recognition APIs. Determine whether the application needs detailed probabilities and assess repeated-query behaviour. Measure the protection provided by restrictions rather than describing one measure as complete anonymisation. Evaluate access to the service as well as the handling of its training files.