Project R-15388

AI & Video Streaming (Research)

This doctoral thesis investigates the impact of Artificial Intelligence (AI) techniques on streaming in general, but is primarily focused on video streaming. Terms like AI and Machine Learning (ML) have suddenly become inseparable form our daily environment, and have a very broad impact on technology and society. The impact of these technologies on video streaming can be viewed at different levels. For example, attempts can be made to make better predictions of the next video frame in a sequence using artificial neural networks, thereby reducing the number of remaining differences that need to be encoded, resulting in a more efficient bitstream. Even in this aspect, ML can be deployed in various ways: only for predicting the next frame (exploiting temporal information), potentially also for encoding reference frames (exploiting spatial correlations), or even by looking at the entire video sequence and having the trained weights of the neural network represent the whole video. Depending on the approach, this can be used for live or pre-recorded video, or both. When it comes to pre-recorded video sequences, as used by many available streaming services, there is typically switching between different qualities for short segments of the video, to offer the end-user an optimal viewing experience based on the (sometimes changing) available bandwidth. Making these decisions is far from trivial, as it can never be known exactly how the transmission network will behave for a future fragment. AI systems might be able to develop better general strategies based on past behaviors, possibly even tailored to individual end-users. Lastly there is also the important security aspect: now that more and more online meeting are taking place, how can one be sure they are dealing with the expected person? How can video be provided with a sort of signature, or can the authenticity of a video stream be guaranteed in another way? Of course, detecting malicious video streams will always be challenged by other AI systems attempting to circumvent such detection.

01 November 2024 - 31 October 2028