Tackle Deepfakes with CoderPush’s On-Device AI Solution
Deepfakes are no longer a future risk. They already threaten financial services, identity verification, and media trust. CoderPush partnered with Ensign InfoSecurity to make detection fast, private, and usable directly on device.
Deepfakes turn trust into an attack surface.
AI-generated videos, images, and audio are now convincing enough to drive financial fraud, identity theft, and reputational damage. A 2024 CNN World report described a finance employee being deceived into transferring $25 million after a deepfake video impersonated a CFO.
Traditional cloud detection can be slow, expensive, and exposed to connectivity limits. That is a real problem in regions where users depend on older devices or unstable networks, and where media verification still needs to happen immediately.
Move the model closer to the moment of risk.
CoderPush and Ensign built an on-device deepfake detector with Google's LiteRT and AI Edge SDK. Instead of sending every recording to the cloud, the system analyzes media locally, cutting latency while improving privacy.
The result was showcased at Google IO 2025 and reduced detection time from 67 seconds to 1.2 seconds. That difference changes the product from a background check into something users can rely on during real interactions.
67s to 1.2s
Detection time dropped enough to make real-time deepfake review practical on device.
Offline first
Processing happens locally, which reduces cloud latency and keeps sensitive media closer to the user.
Older devices
LiteRT, AI Edge SDK, quantization, and CPU/GPU delegates helped support common mobile hardware in Asia.
The app stays simple while the extension does the hard work.
The mobile app is bundled as two coordinated components. Users get a clear recording flow, detection history, and readable results. The extension handles preprocessing, model execution, secure result storage, and upload behavior behind the scenes.
- Main app: recording controls, detection history, results, and ReplayKit Broadcast Picker integration.
- Extension: frame preprocessing, AI model execution, encrypted local result storage, S3 upload, and local completion notifications.
- Security boundary: heavy processing stays separated from the user interface so the app remains usable and protected.
On-device AI has to solve product and systems constraints together.
The team had to balance model accuracy, CPU/GPU delegation, memory use, privacy, and a user experience simple enough for high-stakes moments. Quantization and careful runtime choices helped keep the detector efficient on common hardware.
That work matters because a security tool fails if it is either too slow to use or too heavy for the devices that need it most.
The same architecture can protect more than one workflow.
Financial security
Banks and fintech teams can verify identities in real time and reduce fraud exposure.
Broad accessibility
The model can support lower-end devices and low-connectivity environments common across diverse Asian markets.
Energy efficiency
Lower memory and power usage reduces heat while extending practical hardware lifespan.
Reusable pattern
The same on-device approach can support media authentication, identity verification, image recognition, and IoT security.
Deepfake defense needs product speed, not just model accuracy.
Deepfake attacks are rising quickly, including voice phishing and manipulated video. CoderPush's collaboration with Ensign shows how AI teams can turn model capability into a usable, privacy-conscious security product.
If your organization needs to verify media, protect identity workflows, or bring AI inference closer to the edge, the right next step is to scope the threat model and build the smallest useful detection loop.