The present paper reports on a theoretical and experimental stu dy on the possibility of achieving content-based MPEG-4 AVC video a uthentication. The authenticati on signature is extracted so as to both maximize the probability of correct decision and the mutua l information under content-preserving attacks. The experiments (carried out on 1h of video surveillance corpus) demonstrate that the content-modification attacks can be identified with Precision and Recall rates larger than 0.9 while ensuring a 9s temporal accuracy and an 1/81 frame size spatial accuracy.