UC Irvine

Visual media plays a central role in journalism, public safety, legal evidence, and everyday communication. Yet advances in deepfakes, generative AI, and media editing have eroded trust in what cameras record. This dissertation argues that restoring trust in visual content requires moving beyond post-hoc detection toward preventative, system-level authenticity guarantees that span capture, processing, and consumption. It presents three systems that address complementary weaknesses in modern visual media pipelines.First, Vronicle provides fine-grained provenance for videos captured on mobile devices. By combining trusted execution environments (TEEs) with cryptographic provenance and fixed-function post-processing units, it enables a consumer to verify both the source camera and the exact sequence of post-processing filters applied to a video. Second, ProvCam strengthens the root of trust (RoT) at capture time. It introduces a camera module with a small self-contained trusted computing base (TCB) that protects video capture, image processing, encoding, and signing against software and hardware tampering, thereby reducing reliance on large, attack-prone device stacks. Third, Scoop exposes a fundamental blind spot in provenance-based media authentication: recapture attacks, where fake content is displayed on a screen or printed surface and then captured by a secure camera. Scoop detects misleading recaptures by comparing depth sensed by time-of-flight hardware with perceived depth inferred from RGB images.Together, these systems show that trustworthy visual media requires both strong provenance guarantees and defenses against physical-world attacks that exploit the camera itself. The dissertation demonstrates that these goals are practical on commodity mobile devices and prototype hardware, and that they can substantially improve the authenticity, integrity, and credibility of photos and videos in a world increasingly shaped by synthetic media.