Scam.ai launches on-device deepfake detection for video calls

Not having to worry if someone is real on a video-call on the device most people carry around with them would be reassuring. Real-time inference is making this happen, as Scam.ai and Qualcomm have developed on-device deepfake detection.

Scam.ai worked closely with Qualcomm to develop what the Canadian company calls Halo, a real-time and on-device deepfake detection model for video meetings, which has now launched and with multiple platforms in the release pipeline.

“We didn’t want to build another tool that tells you about a deepfake after the damage is already done,” says Simiao (Ben) Ren, cofounder and CEO at Scam.ai.

Real‑time inference is what happens when an AI system receives real‑world input like text, video or sensor data and immediately generates a result. This could be a chatbot reply, a fraud‑detection alert or a recommendation. The model uses its learned parameters to interpret new data and to run the model locally requires optimization for the best results.

“Working with Qualcomm let us run real-time detection directly on the Snapdragon NPU,” Ren explains, “so Halo can catch a synthetic face the moment it appears on a call, without the video ever leaving the device. That’s the difference between a warning and a report.”

Halo operates fully on the user’s device scanning live Zoom, Microsoft Teams and Google Meet calls for signs of a synthetic face. It issues alerts instantly based on what appears in the meeting. The system runs on Qualcomm’s on‑device AI processor, the Snapdragon NPU, analyzing each video frame locally as the call unfolds. Audio or video never leaves the device for cloud analysis.

Halo is currently available for Windows, with more platforms to follow. New deepfake detection product launches are meeting the rising challenge as the market grows for such solutions, some of which are prototypes.

Scam.ai beefs up its deepfake analysis with voice detection

Modulate and Scam.ai have formed a partnership to bring synthetic voice detection into Scam.ai’s deepfake analysis platform. Customers will be able to screen audio, images and video through a single workflow.

Scam.ai already analyzes manipulated images, videos and digital documents. Now that Modulate’s voice‑detection models are integrated, the platform can flag AI‑generated or cloned audio in both live and prerecorded interactions.

The unified approach handles how scams are moving across multiple channels, starting with a cloned voice before shifting to fabricated documents and ending with manipulated video, the companies say.

“Scammers stopped limiting themselves to one channel a long time ago, but many detection systems are still organized around individual media formats” claims Ren, Scam.ai’s co‑founder and CEO.

Modulate CTO Carter Huffman says voice is becoming a key component of coordinated multimedia scams, often used to create urgency or trust before other synthetic content reinforces the deception.

Modulate’s model, which leads the Hugging Face Speech Deepfake Detection Leaderboard, will be offered directly through Scam.ai’s interface. Scam.ai’s platform already reports high accuracy for detecting manipulated images and video.

Customers will be able to screen synthetic content across all three media types, prioritize high‑risk cases and integrate voice detection into existing fraud prevention and identity verification workflows. Applications range from onboarding and payment authorization to impersonation detection, content moderation and digital evidence review.

Scam.ai expects integrated voice detection to be available in early September. The Canadian startup is in a research partnership, that includes software integration, with fellow Canadian startup Deepidv, which raised a million dollars in seed funding in March.

Article Topics

AI fraud  |  deepfake detection  |  deepfakes  |  Modulate  |  Qualcomm  |  Scam.ai  |  videoconferencing  |  voice biometrics