Deepfakes & the New Identity Threat Landscape

A deepfake is synthetic media — video, audio, or imagery — generated or manipulated by a machine-learning model so that it depicts a real person doing or saying something they never did. The underlying techniques are not new; what changed is accessibility. Face-swapping, lip-sync, and voice cloning that once demanded a research lab and hours of reference footage now run from consumer tools with only seconds of a target's audio. That collapse in cost is why deepfakes stopped being a novelty and became a business risk — most acutely for the identity checks that underpin authentication, payments, and executive trust.

How synthetic identity attacks actually land

The most damaging cases are not viral hoaxes; they are targeted, low-volume attacks against a single high-value decision. A cloned executive voice instructs a finance officer to release an urgent payment. A live video call — with a real-time face-swap of a senior leader — pressures a junior staff member to bypass a control "just this once." A synthetic face is presented to a remote onboarding flow to open an account under a stolen identity. In each case the deepfake does not need to be flawless; it only needs to survive the few seconds of human or automated scrutiny that stand between the attacker and the action.

Why our instincts fail

Humans are poor deepfake detectors. We anchor on voice familiarity and facial gestalt, exactly the signals generative models reproduce best. Worse, the social-engineering wrapper — urgency, authority, secrecy — is engineered to suppress the very skepticism that might catch an artefact. Treating "I recognised their voice" as proof of identity is now a control failure, not a reasonable judgement.

Where organisations get hit hardest

Three surfaces absorb most of the damage. The first is finance and payment operations, where a cloned voice or a live video impersonation of an authority figure short-circuits approval controls. The second is remote identity onboarding — account opening, KYC, and remote proofing flows that rely on a selfie or a short video to bind a real person to a new credential; a synthetic face injected into that flow lets an attacker onboard under a stolen or fabricated identity at scale. The third is the help desk and account-recovery path, historically the softest link in identity, where an attacker who sounds like the account holder can talk their way past a reset. Mapping which of your workflows trust a face or a voice — and treating each as a candidate for compromise — is the first practical step.

Detection: artefacts, but a moving target

Technical detection looks for the fingerprints a generator leaves behind: inconsistent lighting and reflections, unnatural blink and micro-expression timing, blending seams around the hairline and jaw, frequency-domain artefacts invisible to the eye, and physiological signals like the subtle colour shifts of a real pulse that many models fail to reproduce. These detectors are genuinely useful, but they live in an arms race — every generator improvement erodes yesterday's tell. Detection should therefore be treated as one weighted signal in a broader decision, never as a single gate that a "clean" score can open.

Liveness: proving a real person is present

Liveness detection asks a different, more durable question: not "is this media fake?" but "is a live human physically present right now?" Passive liveness analyses depth, texture, and micro-movement from a normal capture without asking the user to do anything. Active liveness issues an unpredictable challenge — turn your head, follow a moving prompt, respond to a randomised code — that a pre-rendered or injected video cannot satisfy in real time. The strongest deployments also defend the capture path itself, because a determined attacker will try to bypass the camera and inject a synthetic stream straight into the application. Certification against standards such as ISO/IEC 30107 for presentation-attack detection gives a defensible baseline.

Detection asks whether the pixels are fake. Liveness asks whether a person is really there. Provenance asks where the media came from. You want all three — no single one is sufficient.

Provenance: signing the truth at the source

Chasing fakes is a losing game if you only ever inspect the output. Provenance flips the model: instead of proving media is fake, prove which media is authentic. The C2PA standard — Content Credentials — attaches a cryptographically signed manifest to media at the point of capture or creation, recording the device, the edits applied, and a tamper-evident chain of custody. Any later modification breaks the signature. Provenance will not stop a determined forger from producing unsigned content, but it lets trusted sources make verifiable claims, so that "no credential" becomes a meaningful warning rather than the silent default.

Rebuilding identity verification

The strategic response is to stop treating a face or a voice as an authenticator. Bind sensitive actions to something an attacker cannot clone in the moment: phishing-resistant, hardware-backed credentials such as passkeys and FIDO2 security keys, which prove possession of a device rather than resemblance to a person. Re-architect high-risk workflows — payment changes, credential resets, executive instructions — around out-of-band verification through a separate, pre-established channel and a callback to a known-good number. And enforce policy that no urgent request, however convincing the voice or face, overrides the control. The technology is only half the fix; the process discipline is the other half.