On-device deepfake detection is now a reality for everyday video calls. Scam.ai, a Canadian cybersecurity startup, has partnered with Qualcomm to launch Halo, a real-time detection system that identifies synthetic faces during Zoom, Microsoft Teams, and Google Meet calls without ever uploading your video to the cloud. The system runs directly on Qualcomm's Snapdragon Neural Processing Unit (NPU), a specialized chip designed to handle artificial intelligence tasks locally on your device.

Why Does On-Device Detection Matter for Deepfake Prevention?

The timing of detection makes all the difference in fraud prevention. Traditional deepfake detection systems analyze video after the call ends or after suspicious content has already been shared, leaving victims vulnerable to scams. Halo changes this equation by analyzing each video frame as it arrives, issuing alerts the moment a synthetic face appears on screen. This means you get a warning before you've been deceived, not a report after the damage is done.

"We didn't want to build another tool that tells you about a deepfake after the damage is already done," said Simiao (Ben) Ren, cofounder and CEO at Scam.ai. "Working with Qualcomm let us run real-time detection directly on the Snapdragon NPU, 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."

Simiao (Ben) Ren, Cofounder and CEO at Scam.ai

Real-time inference, the technical process behind Halo, is when an AI system receives live input like video or audio and immediately generates a result. Running this inference locally on your device rather than sending data to cloud servers offers two critical advantages: speed and privacy. Your video frames never leave your computer, eliminating the latency of uploading to distant servers and the privacy risk of storing sensitive call data in the cloud.

How Does Halo Detect Deepfakes in Real Time?

Halo operates by scanning live video feeds frame by frame, looking for telltale signs of synthetic or manipulated faces. The system uses machine learning models optimized specifically for the Snapdragon NPU, allowing it to process video efficiently without draining your device's battery or slowing down your call. Currently available for Windows, Halo is designed to integrate seamlessly into your existing video conferencing workflow, issuing instant alerts when it detects suspicious activity.

The launch of Halo reflects a broader shift in how the cybersecurity industry is approaching AI-powered fraud detection. Rather than relying on centralized cloud infrastructure, companies are increasingly moving detection capabilities to the edge, meaning directly onto user devices. This approach reduces latency, improves privacy, and makes detection systems more resilient to network outages.

What Threats Does Scam.ai's Platform Now Cover?

Scam.ai is expanding its fraud detection capabilities beyond video. The company has partnered with Modulate, a voice-detection specialist, to integrate synthetic voice detection into its platform. This unified approach recognizes that modern scams rarely rely on a single medium. Instead, fraudsters often orchestrate multi-channel attacks, starting with a cloned voice to establish urgency or trust, then reinforcing the deception with fabricated documents and manipulated video.

• Video Detection: Identifies manipulated or synthetic faces in live and recorded video content across major video conferencing platforms.
• Voice Detection: Flags AI-generated or cloned audio in both live calls and prerecorded interactions, using Modulate's model which leads the Hugging Face Speech Deepfake Detection Leaderboard.
• Image and Document Analysis: Scam.ai already analyzes manipulated images and digital documents, allowing customers to screen synthetic content across all three media types through a single interface.

Modulate's voice detection model is being integrated directly into Scam.ai's platform and is expected to be available in early September. This integration allows organizations to prioritize high-risk cases and incorporate voice detection into existing fraud prevention, identity verification, and content moderation workflows.

"Scammers stopped limiting themselves to one channel a long time ago, but many detection systems are still organized around individual media formats," said Simiao (Ben) Ren, Scam.ai's cofounder and CEO.

Simiao (Ben) Ren, Cofounder and CEO at Scam.ai

How to Protect Yourself From Deepfake Scams in Video Calls

• Enable Real-Time Detection: Use tools like Halo that analyze video frames as they arrive, giving you immediate alerts rather than warnings after a call has ended.
• Verify Caller Identity Through Multiple Channels: If someone claims to be a trusted contact, independently verify their identity using a phone number or email address you already have on file, not contact information provided during the suspicious call.
• Stay Alert to Multi-Channel Attacks: Be aware that scammers often use cloned voices first to establish urgency, then follow up with fabricated documents or video. Skepticism across all communication channels is essential.
• Keep Software Updated: Ensure your device and video conferencing applications are running the latest versions, which include the newest security patches and detection capabilities.

The emergence of on-device deepfake detection represents a significant step forward in the ongoing arms race between fraudsters and security researchers. By moving detection from the cloud to your device, companies like Scam.ai are making it possible to catch synthetic content in real time, before it can be weaponized against you. As deepfake technology becomes more sophisticated, this shift toward local, instantaneous detection may become essential infrastructure for anyone relying on video calls for sensitive conversations.