SPEAKER_1: Last time we established that multi-modal fusion increases privacy risk — more sensors, harder to assess exposure. So now I want to get into what responsible design actually looks like. SPEAKER_2: Right, and the starting point is legal. The EU's GDPR defines biometric data as data produced by specific technical processing of physical or behavioral characteristics that allow unique identification. Gait falls squarely inside that definition. SPEAKER_1: So the legal protection kicks in because of what the processing does — not whether a face appears in the frame. SPEAKER_2: Exactly. A system creates serious privacy risk even without recording a recognizable face. Motion patterns themselves preserve identity across recordings and databases. Removing names from gait records doesn't make them anonymous — the features stay linkable. SPEAKER_1: Wait — de-identification fails even when the data looks stripped down? SPEAKER_2: [short pause] Yes. A study found that realistic full-body anonymization removed visual appearance while preserving enough gait information for identity discrimination. Blurring faces or replacing clothing is not sufficient. The motion itself carries the identity. SPEAKER_1: So what does GDPR actually require designers to do differently? SPEAKER_2: GDPR Article 25 requires data protection to be considered when processing means are selected — not added after the system is built. It names pseudonymization and data minimization as implementation measures. Privacy by default means personal data necessary for a specific purpose gets processed by default, with limits on collection, processing, storage duration, and accessibility. SPEAKER_1: Mm-hmm. And NIST has a parallel framework for this? SPEAKER_2: The NIST Privacy Framework organizes risk management around five functions: Identify-P, Govern-P, Control-P, Communicate-P, and Protect-P. NIST describes privacy risk as the possibility that processing creates problems for individuals — embarrassment, discrimination, economic loss, or loss of trust. Mapping the full data lifecycle, from sensing through disposal, is how a team actually finds where those risks arise. SPEAKER_1: So one practical tool is edge processing — keep the raw data local, transmit a narrowly tailored decision or protected feature. SPEAKER_2: Keeping raw video or inertial data on a local device and transmitting a narrowly tailored decision or protected feature can reduce exposure compared with centralized raw-data collection. The raw video or inertial data is kept on the device. And ordinary encryption doesn't solve it alone — an authorized service can still learn information after decryption. Encrypted-domain inference is a separate design objective. SPEAKER_1: That's the counterintuitive part — the classifier runs on ciphertext and still produces a result? SPEAKER_2: A 2026 study reported gait identification from synchronously sensed mechanical and bioelectrical signals while processing encrypted inputs directly in the ciphertext domain. [chuckle] The math does the heavy lifting. And GaitPrivacyON used an autoencoder to transform mobile gait data into a representation designed to suppress sensitive attributes while retaining authentication utility — achieving area-under-the-curve values above 99% for verification. SPEAKER_1: So privacy protection and high verification utility can coexist. But the evaluation has to go beyond intended accuracy. SPEAKER_2: Exactly — the key idea. Privacy-preserving representations should be tested for unintended leakage of identity, sex, activity, health status, and other attributes, rather than treating recognition performance as the whole evaluation. And biometric templates require special protection because a gait pattern is tied to a person's body and cannot be replaced like a password after compromise. SPEAKER_1: NIST is specific about biometrics as a sole authentication factor, right? SPEAKER_2: NIST states that biometric characteristics are not secrets — they can potentially be obtained without the subject's cooperation. Current digital-identity guidance requires biometrics as part of multi-factor authentication with a physical authenticator, and requires a non-biometric alternative for subscribers. Biometric samples should be erased immediately after any permitted template adaptation. SPEAKER_1: And the FTC has weighed in on the consumer side. SPEAKER_2: Significantly. The FTC warns that biometric technologies create privacy and data-security risks, bias, and discrimination — and expects companies to address foreseeable harms. Surreptitious collection, inadequate risk assessment, weak vendor oversight, and unsupported accuracy claims are all flagged. And here's the one that surprises people: the FTC specifically warns that biometric information can be used to create deepfakes for fraud or harassment. Privacy design must consider downstream misuse — beyond unauthorized identification. SPEAKER_1: The takeaway for everyone following along — privacy isn't a public-relations statement. It's a technical requirement with legal teeth. SPEAKER_2: That's exactly it. Gait processed for unique identification is specially protected biometric data under GDPR, and anonymizing visual appearance is insufficient. Responsible design means purpose limitation before collection, edge processing to minimize transmission, encrypted-domain inference where possible, template protection, and audit logs with deletion policies. Privacy by design is the architecture — not the disclaimer.