Strategies for Countering Deepfakes in Video Conferences through Digital-Identity Assurance

Abstract

Recent developments in deep learning algorithms now empower malicious actors to produce photorealistic, real-time deepfakes that can bypass visual trust cues and compromise established multi-factor authentication (MFA) protocols. This thesis explores how digital identity systems can be rethought to withstand such attacks while remaining practical in enterprise video-conference settings. Building on an empirical assessment of recent deepfakeenabled fraud, the work develops a threat model that highlights the reliance of traditional onetime passwords and push-based MFA on easily manipulated human verification steps. A solution to these challenges is an approach to hardware-based authentication that unifies the factors of possession, inherence, and context in a single secure transaction. The new prototype two-factor authentication device includes a display, a secure element, a capacitive fingerprint sensor, and a GPS module. An Arduino platform leveraging these four components produces and displays a QR token valid for 30 seconds with content secured by the Advanced Encryption Standard (AES-256) encryption and authenticated using the Hash-based Message Authentication Code - Secure Hash Algorithm (HMAC-SHA). It is designed based on the principles of confidentiality, integrity, availability, and authenticity in NIST SP 800-63-3, as well as key Common-Criteria assurance categories for user authentication and fail-secure operation. Operational tests indicate verification times of approximately 5 seconds, from scanning to displaying the result. The device also features resistance to protocol replay, manipulation, and MFA-fatigue attacks that have recently affected high-profile organizations. The findings demonstrate that carefully integrating hardware trust anchors in organizations with compact, QR-code-mediated (air-gapped) cryptography can help restore confidence in remote identity verification and make compromise more difficult, even as synthetic media capabilities continue to improve.

Date of Award	2025
Original language	English
Supervisor	Johannes Edler (Supervisor)