Secure Voting System Using Anti-Spoof Face Recognition and Deep Learning with Encryption

Identification and prevention of fraud are still the biggest challenges in electronic voting systems. Here, we propose a multi-layer Secure Voting System that utilizes face recognition, along with deep learning-based spoofing detection, face comparison-based authentication, and secure vote encryption, to address these issues. Eligible voters are provided with a unique National Identity Card (NID) that contains biometric information, including photographs and fingerprints. The NID number serves as a crucial authenticator for voter verification. The system workflow begins with requesting the user to enter his/her NID number and capturing a live image through the webcam. Liveness detection is performed using a Convolutional Neural Network (CNN) model trained on a publicly available anti-spoofing dataset comprising $\text{3 0, 0 0 0}$ samples. The model can distinguish between live faces and spoof attempts using printed images, cutouts, and video replays in real-time. Our system employs a multi-layer model architecture, where after successful spoof detection, the input is passed to a separate face-matching model trained on a dedicated dataset containing 1,998 training images and 114 test images across five identity classes. This layered approach enhances verification robustness by combining deep learning models for sequential validation. The face-matching model identifies the user and verifies their identity by cross-matching the live face with the biometric information related to the input NID. If the face ID and NID do not match the stored records, the system blocks access, displays a message saying try again later, and directs the user to the homepage. If the verification is successful, the voting system enables the user to cast their vote. Once the vote has been cast, it is encrypted instantly and stored securely to ensure data integrity and voter anonymity. To prevent multiple votes, the user's NID is removed from the eligible list upon a successful vote, ensuring that only a single vote can be cast per individual. Role-based access control is implemented, enabling access only to administrators of results while voters remain completely private. Experiment performance is demonstrated to be 96 % accuracy and weighted F1 of 0.95 for the anti-spoofing model, with excellent performance even against class imbalance. The face matching model achieved 89 % accuracy, with perfect classification for three out of five identity classes and a macro F1 score of 0.89, reflecting stable and reliable identity verification across users. The integration of multi-layer face verification and secure encryption offers a highly secure, tamper-proof electronic voting system.