Scam.ai and Qualcomm Bring On-Device Deepfake Detection to Video Calls

At Computex 2026 in Taipei, Scam.ai unveiled Halo, an on-device deepfake detection model for live video calls, alongside a strategic partnership with Qualcomm. This move marks a significant shift from cloud-based detection to real-time, edge-based inference, addressing the growing threat of synthetic media in enterprise communications.

Related Articles

• OpenAI turns AI loose on open source bug hunting

• Mythos 5 Goes Government: What 100+ US Agencies Means for AI

• Scam.ai’s Halo model runs entirely on-device, enabling real-time deepfake detection during live video calls without cloud latency.

• The Qualcomm partnership brings Halo to desktop platforms, leveraging Qualcomm’s AI Engine for efficient inference.

• On-device processing addresses privacy concerns by keeping video data local, reducing data breach risks.

• This launch comes as deepfake incidents in corporate video calls have surged, with a 2025 report noting a 300% year-over-year increase in such attacks.

• Halo’s deployment at Computex 2026 signals a growing trend of edge AI in cybersecurity, moving beyond traditional server-based solutions.

• Enterprise adoption of on-device AI for security is expected to accelerate, driven by regulatory pressures and cost savings on cloud compute.

How Does Halo Detect Deepfakes in Real Time?

Halo operates by analyzing video frames locally on the user’s device, using a lightweight neural network optimized for Qualcomm’s hardware. The model looks for subtle artifacts in facial movements, lighting inconsistencies, and audio-visual mismatches that often betray synthetic media. By processing at the edge, latency drops to under 50 milliseconds, making it viable for live conversations. This approach contrasts with cloud-based detectors that introduce delays and require constant internet connectivity. Scam.ai claims Halo can flag manipulated feeds before a human would notice anything amiss, a critical feature for executive briefings and board meetings where trust is paramount.

For organizations deploying video call security, prioritize models that offer per-frame analysis with confidence scores. This allows security teams to set thresholds for alerts, reducing false positives during high-stakes negotiations.

Why Is On-Device Detection Harder to Get Right Than It Looks?

Building an on-device deepfake detector presents unique challenges. The model must be small enough to run on consumer hardware without draining battery or overheating, yet accurate enough to catch sophisticated forgeries. Qualcomm’s AI Engine provides tensor acceleration, but Scam.ai had to prune Halo’s architecture to fit within strict compute budgets. Another hurdle is dataset diversity: training data must cover varied lighting conditions, camera qualities, and ethnicities to avoid bias. Early tests showed that models trained on high-quality deepfakes struggled with low-resolution webcam feeds, necessitating synthetic data augmentation. The partnership allows Scam.ai to test on Qualcomm’s reference designs, ensuring broad compatibility across laptops and mini PCs.

Aspect	Cloud-Based Detection	On-Device Detection (Halo)	Impact
Latency	200-500 ms (network round trip)	<50 ms (local inference)	Real-time alerts in live calls
Privacy	Video stream sent to cloud servers	Data stays on device	Reduced exposure, GDPR compliance
Cost	Per-minute cloud compute fees	One-time hardware cost	Lower operational expense for enterprises
Accuracy	High (large models)	Comparable (optimized model)	Trade-off between size and precision
Internet Dependency	Required	Optional (offline capable)	Works in low-connectivity environments

What Should Enterprise Teams Know Before Adopting Halo?

Deploying on-device deepfake detection requires careful integration with existing video conferencing platforms. Scam.ai provides SDKs for major operating systems, but IT teams must ensure Halo is updated regularly to counter new deepfake generation techniques. A pilot rollout on a subset of executive devices is recommended before full deployment. According to the NeuralPress AI Statistics & Trends 2026 resource, enterprise AI adoption reached 78% in 2026, yet only 12% of firms have dedicated deepfake countermeasures, highlighting a significant gap. Halo addresses this gap but requires staff training to interpret alerts and avoid alarm fatigue.

Who Benefits Most From This Development?

• Corporate executives and board members: Regular targets of deepfake impersonation in video calls, they gain a safety net against social engineering.
• Financial services firms: With high-value transactions conducted via video, real-time verification reduces fraud risk.
• Legal and compliance teams: On-device processing helps meet data residency requirements, avoiding cloud data transfer issues.
• Remote work organizations: As hybrid meetings become standard, Halo provides a consistent security layer across diverse hardware.

No detection model is foolproof. Halo may miss advanced deepfakes that leverage real-time generative adversarial networks (GANs). Enterprises should combine automated detection with human verification protocols, such as pre-agreed code words.

Which Warning Signs Predict Problems Ahead?

Adoption of on-device detection will likely face friction from several angles. First, device fragmentation: Halo must support a wide range of Qualcomm Snapdragon and other chipsets, each with different AI capabilities. Second, model update logistics: pushing frequent updates to thousands of endpoints without disrupting workflows is non-trivial. Third, adversarial adaptation: as Halo improves, deepfake creators will evolve their techniques, triggering an arms race. Scam.ai’s roadmap includes federated learning to improve the model without centralizing data, but this introduces its own complexities around model poisoning.

Looking ahead, the partnership between Scam.ai and Qualcomm could set a precedent for other security vendors to embed AI at the edge. As deepfake technology becomes more accessible, the line between authentic and synthetic communication will blur further. On-device detection offers a pragmatic response, but its long-term success hinges on continuous innovation and industry-wide standards for verification. The Computex 2026 announcement is a critical step, but the real test will come when Halo faces real-world attacks at scale.

Source: AI News