War_Crimes_Detection

War Crimes Detection in Videos Project

Yuval Levi and Maxim Katz

Automated detection of war crimes in videos is a groundbreaking new topic that entered our lives during the various wars in the world. In war crime cases, the crime happens between soldiers and civilians when the soldiers use excessive force against the civilians or civilian objects. With the increase in videos documenting war crimes, it was necessary to use artificial intelligence tools to identify these war crimes and warn that a video contains a war crime. There are many purposes for this, first, security officials can use it to collect war crime videos from a wide variety of social networks, for example, in Israel, many cyber security companies are collecting digital evidence of war crimes to base a case for the crimes Hamas terrorists committed[17]. Second, for anxious private individuals who do not want to see war crimes, the war crimes videos can be removed from social networks. We used two architectures, first, Convolutional Neural Networks(CNNs) and InceptionV3 which achieved the highest accuracy in each of the five experiments we committed, and specifically, the highest accuracy and AUC achieved on the Sports Violence dataset with 98% and 99% respectively. For the second architecture, we used Recurrent Neural Networks(RNNs) such as BiLSTM to detect war crimes. This architecture was tested on the same five experiments the first architecture was tested on and achieved its best accuracy of 88% and AUC score of 87% on the Sports Violence dataset. The positive database was the information from 313 videos we collected from the Telegram social network. The negative databases were violent videos from different life situations like violence in sports, fights in the street, and other videos that show soldiers in uniform in various military activities.