SecureSnap

Product Builder · Adversarial ML · On-device AI

Full adversarial ML pipeline — U-Net perturbation generator + PGD server-side cloaking, exported to TFLite and woven into a real Android product that protects photos from deepfake misuse.

APK soon

Status: Working · adversarial ML in-product

Train

Inference

Output

Model pipeline

Research → result

Lab-style panels for training, inference, and outputs.

System design

Overview

An Android product that applies on-device image protection to reduce misuse of photos for deepfake / face-swap generation — while keeping the photo looking normal to humans.

What I built

U-Net perturbation generator trained for photo cloaking
PGD-inspired server-side / training cloaking engine
TFLite export (`protector_net.tflite`) running on-device in Android
Protect → preview → share flow designed for real photo use
Firebase Auth/analytics + remote feature control
React admin panel for operations and monitoring
Model-training repo with CelebA pipelines and evaluation scripts

What's unique

Not a notebook demo: U-Net protector trained, exported to TFLite, and shipped inside the app
PGD-style cloaking thinking on the training/server side, then on-device inference for real users
Offline-first protection with optional Firebase analytics, remote control, and admin panel

Why it matters

Shows production ML craft — train → export → integrate → ship
Privacy-preserving photo sharing without destroying visual quality
Recruiters see adversarial ML applied to a consumer Android workflow

Craft

How the intelligence was built

PythonPyTorchU-NetPGDTFLiteAndroidJavaFirebaseReact

Open the research

Explore SecureSnap

APK soon