Sentin-Edgde

Home Page
Deepfake detected
Real Image Calssified

Inspiration

In an era where generative AI can mimic reality with terrifying precision, the concept of "seeing is believing" is under threat. We noticed a dangerous gap: most deepfake detection happens in the cloud, forcing users to sacrifice their privacy to verify the truth. We were inspired to build Sentin-Edge to give users a "Reality Firewall" that lives entirely on their device—protecting their data while exposing synthetic deception in real-time.

What it does

Sentin-Edge is an on-device forensic suite optimized for the Samsung S25 Ultra. It employs a four-layer "Forensic Firewall" to analyze images and videos:

Safety Layer: Scans raw file URIs for metadata provenance (C2PA/IPTC) and known AI watermarks.

Visual Ensemble: Uses NPU-accelerated Vision Transformers (ViT) and ResNet50 to detect sub-pixel artifacts and generative noise.

Temporal Engine: Tracks facial landmark stability (HRNet) across 20-frame windows to catch the "jitters" common in real-time deepfake overlays.

Forensic Reasoning: Utilizing Gemma 4, the app translates complex technical scores into clear, human-readable explanations.

How we built it

The project is built on the Snapdragon 8 Elite platform using the Qualcomm AI Stack. We utilized LiteRT (formerly TensorFlow Lite) to deploy a multi-model pipeline.

We leveraged the CompiledModel API for AOT (Ahead-of-Time) compilation, ensuring our vision models hit the Hexagon NPU with minimal latency.

We integrated LiteRT-LM to run the 2GB Gemma 4 E2B model with NPU acceleration, enabling agentic reasoning without ever leaving the device.

The UI was built with Jetpack Compose, designed to provide a "security-first" dashboard experience.

Challenges we ran into

The primary challenge was Memory Orchestration. Running a 2GB LLM alongside high-fidelity vision models like HRNet and ViT pushed the mobile RAM limits. We overcame this by implementing a staggered initialization strategy and using 8-bit quantization (w8a8) for all models to maximize NPU efficiency. Another hurdle was preserving metadata during analysis; we solved this by implementing a Direct-URI access system that interrogates the raw file system before standard Android media compression occurs.

What we learned

We learned the true power of On-Device AI. Moving from CPU-based inference to the NPU didn't just make the app faster—it enabled a new category of "Privacy-Preserving Forensics" that wasn't possible before. We also discovered that "Explainable AI" is just as important as the detection itself; giving a user a reason is far more effective than just giving them a score.

What's next for Sentin-Edgde

We plan to expand our temporal engine to include audio-visual sync detection (matching lip movements to vocal frequencies) and integrate a system-wide "Screen Guardian" mode that can protect users during live video calls in any application.

APK is available on this github release: https://github.com/namrathavpatil10/google_hackthon/releases/tag/v1