Sponsored Research
Permanent URI for this collectionhttps://hdl.handle.net/20.500.11875/4529
Browse
Browsing Sponsored Research by Subject "deep learning"
Now showing 1 - 2 of 2
- Results Per Page
- Sort Options
Item Deep Learning Approaches for Fingerprint Verification(Institute for Homeland Security, 2023-10-15) Dalvi, Nikita; Pham, Van VungFingerprint verification is vital because it provides a unique and permanent way to identify individuals. This technology is widely used in various areas like law enforcement, access control, and identity verification processes. Existing approaches for fingerprint verification tasks suffer from low accuracy due to training directly on low-quality and latent fingerprints. Therefore, this work proposes to utilize recent advancements in deep learning and computer vision to (1) enhance fingerprint image quality; (2) extract and verify that the minutiae are retained after enhancement; and (3) perform fingerprint verification tasks. Specifically, this work experiments with (1) Super-Resolution Convolutional Neural Network (SRCNN), Fast SRCNN, and Very Deep Super Resolution (VDSR) for fingerprint image enhancement; (2) Finger-Flow for minutia extraction; and (3) Siamese neural network for fingerprint verification. The experiment results indicate that among the experimented super resolution approaches, VDSR outperforms the others. Additionally, it can retain minutiae in the enhanced version and shows great potential to enhance latent fingerprints, which are less visible. Most importantly, the verification performances improve on the enhanced fingerprints versus low-resolution counterparts.Item Detecting Deepfakes under Anti-forensics Attacks(Institute for Homeland Security, 2023-10-15) Liu, Qingzhong; Celebi, Naciye; Zhou, BingWhile AI is vastly evolving, wherein deepfake techniques may be used to generate more realistic faces, voices, and videos, many deepfake-based fraudulent cases are increasingly occurring. To combat deepfake-based forgery, several methods have been proposed wherein the most astonishing methods are based on convolution neural network (CNN). However, most intelligent detection systems are underrepresenting in exposing the deepfake images under anti-forensics attacks, e.g., rescaling the image, inserting noises, and compressing the image again. To our knowledge, it still falls short of an intelligent detection system being able to detect deepfake and other advanced image forgery together. Additionally, it falls short of a comprehensive comparison study on the latest deep learning models for the deepfake detection. In this study, we apply the latest deep learning models for deepfake detection under pos anti-forensics processing mixed with seam-carving and copy-move forgery images in JPEG. Our study shows that different deep learning models have different distinction capability. Experimental results show that some latest deep learning models are effective in detecting deepfake images under post anti-forensics processing in JPEG images, they are also performing well in detecting seam-carving and copy-move forgery. Our study also shows that it is relatively easy to detect deepfake compared to the detection of seam carving forgery detection under antiforensics processing in JPEG images.