Intelligent general evaluation method and system for vulnerability of recommendation system

Abstract

The invention relates to an intelligent general evaluation method and system for the vulnerability of a recommendation system. The method comprises: 1) establishing a search tree of an attack strategyfor a recommendation system, 2) learning an efficient attack strategy from the search tree by adopting a reinforcement learning algorithm, and 3) evaluating the vulnerability of the recommendation system according to the attack strategy obtained by learning and the test attack effect. Aiming at a complex recommendation system, the invention provides a scheme for efficiently and adaptively evaluating the vulnerability of the complex recommendation system. The scheme is based on the reinforcement learning architecture, excessive manual intervention is not needed, the labor cost is low, the effective attack means for the specific recommendation system can be quickly positioned, the vulnerability of the recommendation system is accurately and efficiently evaluated through the attack effect, and the obtained efficient attack strategy can also guide the design of the defense means in a more targeted manner.

Description

technical field [0001] This method belongs to the field of information technology. It is mainly aimed at the current popular online recommendation systems. It considers how to quickly and adaptively evaluate its ability to withstand hacker attacks, that is, its vulnerability, for different recommendation systems in the context of limited knowledge. This method is based on the advanced reinforcement learning technology architecture, without too much manual intervention, can efficiently explore useful attack strategies, and quickly detect the performance of different recommendation systems under different hacking strategies, so as to quickly detect their vulnerabilities. Background technique [0002] In order to better serve users, a large number of online system platforms at home and abroad have introduced more abundant and advanced algorithms to guess the preferences of new and old users, so as to recommend platform content more targetedly. One of the most basic ideas behind...

AI Technical Summary

Problems solved by technology

The former requires the attacker to have strong domain knowledge, artificially inductive attack strategies, and the cost is high, and for the platform itself, it requires a lot of manpower and material support; the latter can achieve good results by using some optimization algorithms without requiring a lot of manpower , but the premise is that the system has a high level of authority, and at the same time, there are strong restrictions on the recommended system being attacked. It can only target a relatively single and simple recommendation algorithm

The current recommendation system, on the one hand, has high privacy, and it is impossible for an attacker to have high authority on the recommendation system. On the other hand, with the development of deep learning and the increase in the complexity of business processes, it is no longer a single simple algorithm. expressive

Therefore, this kind of algorithm shows different attack effects on different recommendation systems, and cannot give convincing vulnerability evaluation results.

Images