Researchers Measure the Effort to Predict an AI-Generated Number
The study of logits from open models shows that a number presented as random can be linked to an exploitable distribution. Converted into bits and attempts, these distributions provide a concrete measure of the level of security and the decryption effort associated with it.
Security and Effort: What the Distributions Reveal
The analysis of response distributions from AI models allows for the evaluation of the security of the generated numbers and the estimation of the effort required to decrypt them. Examining the logits indicates how many attempts would be needed to guess a number produced by the AI.
Le brief IA que les pros lisent chaque soir
Les 7 actus IA du jour, décryptées en 5 min. Gratuit.
Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.
Choisissez votre rythme
Gratuit · Pas de spam · Désabonnement en 1 clic
From Probability to Bits and Attempts
The study relies on the analysis of logits to measure the probability of obtaining an exact response. These probabilities are then converted into bits and the number of attempts necessary to decrypt or guess the generated number, allowing for a concrete quantification of the decryption effort.
Affected Models: Base and Instruction-Tuned
The analysis focuses on open models, whether they are base models or tuned for instruction. In both cases, the observed response distributions can be exploited, raising questions about the truly random nature of the numbers produced by these AIs.
Brief IA — L'actualité IA en français
L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.