I work in the areas of Information Forensics and Security, Visual Surveillance and Steganalysis.
Information Forensics and Security
Media Forensics serves the authentication and source identification purpose. It is useful when the content is not available apriori to apply active security techqniues such as Digital Signatures and Information Hiding. In such cases, the task is to determine the processing history of the media. This includes determining whether the test media has been processed using Image/Signal processing techqniues with any malign intentions. In many cases, multimedia evidences are often produced in a court of law. However, they may not be treated as a primary evidence in the absence of robust forensic tools which can authenticate these evidences. In some other cases, media may be doctored and distributed through social platforms such as Youtube, Facebook and others. One of the recent examples is one of the released JNU videos. In all these scenarios, we need forensic techniques to determine if the given media is authentic or tampered with.
I am also working on the counter-forensics part. The forensic algorithms designed may be easily attacked by smart adversaries. In counter-forensics the aim is to design anti-forensics algorithms to find the robustnes of the forensics algorithms.
In addition to images and videos, audios are also part of forensic investigations. Towards this, I am interested in both audio authenticity determination using Electrical Network Frequency (ENF) and forging audio. ENF signal gets embedded in audio due to nearby electro-magnetic activities. The ENF signal is a unique signature of audio and can also be used for determining the location where audio was captured.
On the other hand, the audio forgery by manipulating ENF signal has received limited attention due to the fact that ENF analysis may easily reveal forgery. However, this can be achieved through careful filter desgining and tracking the ENF signal.
Visual Surveillance
Visual Surveillance can be used for multiple purposes such as traffic monitoring, human re-identification, crowd management, group activity recognition and other civil and military surveillance applications. In particualr, I am interested in investigating optimal appearance and motion models for visual tracking in aerial imagery. These videos can be acquired using drones. In addition to the static camera video tracking challenges such as complex background, occlusion and illumination changes, these videos pose additional problems. The objects of interest may be very small due to low video resolution and also have a huge pose variation. In such a scenario, the popular appearance and motion models lead to poor tracking performance.