Modelling and control of jailbreak attacks in AI systems as hybrid cyber–physical security threats

Résumé fourni par la source

The work elaborates a hybrid cyber-physical architecture to describe and offer solutions to AI jailbreak assaults against safety-oriented frameworks. Our hybrid automaton model of the closed-loop system is a discontinuous system of hybrid mode with discrete choices and continuous dynamics, jail breaks modeled as a coordinated antagonism of cyber inputs, physical disruptions and discrete choices. The policy of the AI is linearized in small adversarial perturbations, and closed-form floods of safety violations when switching between and during modes are obtained. We measure jail break success through spurious jump probabilities, safety barrier deterioration and consider the effectiveness of an attack to be proportional to the AI sensitivity, cross-domain couplings and manipulations of the guard functions. A stochastic dynamical model can be used to describe the dynamics of attack penetration; the measures of resilience being based on steady-state distributions and mean first-passage times. And, lastly, we suggest a detection-based mitigation mechanism based on control barrier functions and MPC-based adaptive defense, in which we guarantee that the systems are in the state of Lyapunov stability with a limited adversarial energy.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.