news

on september 6, at the 2024 bund conference "protecting ai with ai - offense and defense in the era of big models" forum, the "global ai offense and defense challenge", the first domestic technology competition with the theme of big model offense and defense, was launched. the competition focuses on the industrial practice of ai big models, and has designed two-way tracks for offense and defense. it invites white hat hackers and technical talents to conduct actual combat exercises on "data poisoning" of wenshengtu big models, as well as anti-counterfeiting detection competitions for content generated by big models in financial scenarios, to compete for a science and technology prize of one million yuan.

the competition is jointly organized by the chinese society of image and graphics, ant group, and cloud security alliance (csa) greater china, and is co-sponsored by tsinghua university, shanghai jiaotong university, zhejiang university and other c9 universities and a number of industry-university-research organizations, with shanghai artificial intelligence laboratory as a technical partner. the competition aims to gather the strength of academia and industry through technical competitions, face and solve the potential risks in large model applications, and promote the healthy and sustainable development of the global ai industry.

wang yaonan, an academician of the chinese academy of engineering and president of the chinese society of image and graphics, said that in the era of big models, how to ensure the security and controllability of ai systems and how to effectively deal with potential risks in the supply chain have become important issues that need to be solved. this ai attack and defense challenge is a global technology feast jointly created by the chinese society of image and graphics and well-known companies and institutions in the industry such as ant group, focusing on the big model aigc track. from face verification to generative content detection, we look forward to discovering more innovative solutions to further ensure the security of big model applications and industry security.

the competition has two major tracks, "attack" and "defense", which respectively focus on the security of the big model itself and the anti-counterfeiting detection of the content generated by the big model and the risk detection of abuse of the big model, covering test points in multiple algorithm fields such as machine learning, image processing and computer vision, and data processing.

among them, the "attack track" focuses on the actual application risk issues of the large model of raw images. contestants can use a variety of dynamic attack induction techniques such as target hijacking, scenario introduction, and logical nesting to induce the large model to output risk images, thereby activating the potential weaknesses and loopholes of the large model and enhancing the security immunity of the large model raw images. the "defense track" focuses on the credential tampering detection in financial scenarios in ai core to cope with the increasingly severe risks of deepfake and aigc fake certificates. the competition provides a training set of millions of credential tampering data. contestants need to develop and train models, and use the corresponding test set to evaluate the effectiveness of the model and give the probability value of data forgery.

the competition officially started registration on september 6, and the competition review will be completed in early november. from now on, contestants can register through the official website of the chinese society of image and graphics, the official website of alibaba cloud tianchi big data crowd intelligence platform, and other channels. the competition has a prize pool of nearly 1 million yuan to select and commend outstanding talents in the field.

source: beijing daily client

reporter: sun qiru