Scam.ai has partnered with Qualcomm to run its Halo deepfake-detection model locally during live desktop video calls. The product watches a conferencing session in the background and flags suspected synthetic video without sending footage to a cloud service.
The target is a growing identity gap. Traditional controls may verify an account or device, then assume the face and voice on screen belong to the authorised person.
On-device processing protects both speed and privacy
Local inference avoids the network delay of uploading continuous video and keeps sensitive interviews or executive discussions on the computer. Qualcomm is providing optimisation and device-ecosystem support.
Halo is intended to work passively across existing conferencing tools rather than require a new call platform. That can improve adoption, provided users understand when analysis is active and administrators can manage updates.
Human-resources teams face candidates using synthetic identity, while executives may receive calls designed to authorise payments or disclose information. The source cites a dramatic increase in deepfake attempts and low confidence among HR leaders.
Those claims come from a company announcement and need independent validation. Buyers should test Halo across cameras, lighting, compression, skin tones, disabilities and network conditions representative of their workforce.
A detection score cannot become an automatic accusation
False positives carry serious consequences in recruitment and employment. Poor bandwidth, virtual backgrounds or assistive technology can produce unusual visual signals without deception.
Halo should provide a confidence level and enough evidence for a trained reviewer, not label a person fraudulent. Organisations need a neutral verification route, such as a second channel or identity document process, and a way to appeal.
False negatives matter too. Attackers can adapt once they understand detector behaviour, making continuous testing and signed model updates essential.
The system should minimise stored data and protect detection logs. A record connecting a named candidate with a fraud suspicion is sensitive even when no video leaves the device.
On-device execution also changes support. Security teams must know which processor and model version produced an alert, whether the detector was healthy and if local malware could disable it. Signed updates, tamper evidence and central policy can preserve confidence without transferring the underlying call footage.
Employees and applicants must receive clear notice. Covert biometric-style analysis can create legal and trust problems independent of technical accuracy.
Deepfake defence belongs inside a broader identity process
Video authenticity does not prove that a request is legitimate. A real executive account can be compromised, and a genuine person can be coerced. High-value actions still need dual approval and verification through a trusted workflow.
Recruiters should combine call analysis with account history, document checks and consistent interview practice. Detection becomes one signal rather than a new single point of trust.
Security teams need an incident playbook explaining who reviews an alert, how the call continues and when evidence is escalated. A warning that arrives without a response process merely transfers uncertainty to the user.
Performance metrics should include detection by attack type, false-positive rate, review time and outcomes after secondary verification. Vendor claims about real-time operation should be tested under normal device load.
Halo’s local design addresses two genuine obstacles to deepfake detection: latency and reluctance to send confidential calls elsewhere. Its value will depend on whether it remains accurate outside demonstrations and whether organisations use it proportionately.
Trust in a call cannot come from a green indicator alone. On-device detection can raise the right question at the right moment, while accountable people and resilient business controls determine what happens next.
Enterprises should compare Halo with challenge-response procedures and other detectors, then choose layers appropriate to the value of the interaction.
