Skip to content
WITNESS · Physical security sensing

DPIA template. For ReID-via-embeddings.

A Data Protection Impact Assessment structured against the ICO published template, adapted to the specifics of ReID-via-embeddings. The artefact your procurement team needs in the evidence pack before the pilot — fill in the PLACEHOLDER fields, sign Step 7, file with your DPO.

ReID-via-embeddings is not facial recognition. The system extracts pseudonymous 512-dimensional vectors from person detections; no identifying facial features are stored in the embedding, and no images persist beyond the on-camera ring buffer. Step 2 below describes the processing in full.

Site-specific fields.

Operator (data controller)
PLACEHOLDER — replace with your registered company name
ICO registration number
PLACEHOLDER — your ZA reference
Sites in scope
PLACEHOLDER — list each site, address and unit count
Resident / visitor count in scope
PLACEHOLDER — total residents + estimated weekly visitors
DPO (data protection officer)
PLACEHOLDER — name, email, ICO registration if external
SIRO (senior information risk owner)
PLACEHOLDER — name and title
Assessment date
PLACEHOLDER — DD MMM YYYY
Review cadence
PLACEHOLDER — annual by default; sooner on material change

Step 1 — Identify the need for a DPIA.

Outcome. A DPIA is mandatory for this processing. ReID-via-embeddings constitutes systematic monitoring of a publicly accessible area on a large scale, which triggers UK GDPR Art 35(3)(c) automatically. The ICO list of operations requiring a DPIA also names "use of new technologies", which applies to embedding-based ReID even where individual elements (CCTV, similarity search) are mature.

Screening questions answered yes:

  • Systematic monitoring of a publicly accessible place on a large scale.
  • Innovative technological solution (embedding-based ReID, distinct from facial recognition).
  • Processing pseudonymous data that could re-identify individuals when combined with operator-held context.
  • Processing in a context where data subjects (residents, visitors) have a reasonable expectation of safety but limited practical ability to opt out.

Because two or more screening questions return yes, a DPIA is required before processing begins.

Step 2 — Describe the processing.

Nature of the processing. Existing IP cameras on the operator's estate stream to a UK-resident edge node. The edge node runs person detection, extracts a 512-dimensional embedding vector from each detection, and forwards only the embedding (plus a content-addressable id and a timestamp) to the embedding store. Similarity search runs against the embedding store on operator query. No source frames are retained beyond the on-camera ring buffer (default 72 hours; configurable shorter).

What is and is not stored.

  • Stored: 512-dim embedding vectors, content-addressable ids, timestamps, camera ids, event metadata, audit log entries.
  • Not stored beyond on-camera ring buffer: source video frames, bounding-box pixel crops, raw RTSP streams.
  • Not extracted into the embedding at all: identifying facial features (the model is trained on whole-body re-identification, not face recognition), biometric templates as defined by UK GDPR Art 9, or any data that on its own would constitute special-category personal data.

Scope. The processing covers the sites listed in the meta block above. PLACEHOLDER — confirm the estimated count of unique persons detected per day across the estate.

Context. The data subjects are residents, visitors and (incidentally) passers-by in publicly accessible parts of the estate. Residents are informed via the operator's privacy notice and via signage at every camera location. Visitors are informed via signage. There is no contractual relationship with passers-by; lawful basis for their incidental processing is Art 6(1)(f) legitimate interest, balanced in Step 4.

Purposes. The processing exists to enable safety-related operational decisions (incident review, missing-person response, repeat-offender detection across the estate). It is not used for marketing analytics, footfall research, or employee monitoring. Repurposing requires a fresh DPIA.

Data flows. Camera → edge node (UK) → embedding store (UK) → operator UI (UK). No data leaves the UK perimeter. The full topology is documented at /sovereignty.

Step 3 — Consultation.

Internal consultation. The DPO is consulted at every stage of this DPIA. The SIRO signs Step 7. The operations team and the procurement team are consulted on the operational realities of the processing.

Resident consultation. Residents are consulted in advance of deployment via the operator's standard channels (resident association, written notice, dedicated email or town-hall session — PLACEHOLDER — confirm channel). The consultation describes the system in plain English, the lawful basis, the retention windows, and the route to raise concerns. Resident feedback is recorded and material objections are addressed before processing begins.

Visitor consultation. Visitors are not consulted individually but are informed via signage at every camera location and via the operator's published privacy notice. The signage names the controller, the purpose, the retention period and the route to a data subject access request.

ICO consultation. Prior consultation with the ICO under UK GDPR Art 36 is not required for this processing on the residual risk profile recorded in Step 5 below. If the residual-risk band moves to high for any individual risk after Step 6 controls are applied, the DPO will trigger prior consultation before processing begins.

Other consultation. Where the site is covered by a joint controller agreement (e.g. a shared forecourt with a neighbouring building), that controller is consulted and the agreement is recorded in the operator's DPA pack.

Step 4 — Necessity and proportionality.

Lawful basis. UK GDPR Art 6(1)(f) legitimate interest. The legitimate interest is the safety of residents, visitors and staff on the operator's estate. The balancing test below establishes that this interest is not overridden by the rights and freedoms of data subjects, given the controls in Step 6.

Balancing test.

  • Purpose test. The interest is legitimate, real and present — incident response on the operator's estate is a routine and lawful activity.
  • Necessity test. Less intrusive means (rotating human guards, perimeter signage alone) have been considered and rejected as insufficient at the estate scale. ReID-via-embeddings is the least intrusive means of achieving the purpose because no source images persist and no faces are extracted.
  • Balancing test. The data subject's reasonable expectation is that they may be filmed in publicly accessible parts of the estate; the embedding-based approach is materially less intrusive than facial recognition (which is not used) or long-retention CCTV. Risks to the data subject are constrained by the Step 6 controls, particularly retention limits, encryption, audit logging and the human-in-the-loop requirement on consequential actions.

Special-category data. The system is designed to avoid processing special-category personal data under UK GDPR Art 9. Embeddings do not encode identifying facial features and are not biometric templates as defined by Art 4(14). If the operator's deployment context creates a foreseeable path to special-category data (e.g. cameras covering a place of worship or a medical facility), a separate Art 9 assessment is required before deployment.

International transfers. None. All processing is UK-resident. The full topology is documented at /sovereignty.

Step 5 — Identify the risks.

Each risk is rated for likelihood and severity with a documented rationale. The corresponding controls are listed in Step 6 below.

  • R1

    Likelihood LOW

    Severity HIGH

    Risk Embedding reverse-engineering — a captured 512-dim vector is fed back into a public model to recover an approximation of the original face or body.

    Rationale Public embedding-inversion research exists but assumes attacker access to (a) the embedding store and (b) the exact model weights. We hold weights privately and the store is encrypted at rest. Severity is high because inversion would constitute biometric data under UK GDPR Art 9; likelihood is low because the attacker would need to compromise two distinct systems.

  • R2

    Likelihood MEDIUM

    Severity HIGH

    Risk Model bias — the ReID similarity metric performs worse on certain demographic groups, producing differential false-positive rates.

    Rationale ReID models trained on imbalanced datasets are known to underperform on under-represented groups. Severity is high because differential error rates create discriminatory outcomes; likelihood is medium because mitigations exist (balanced training data, per-group calibration) but cannot fully eliminate bias.

  • R3

    Likelihood LOW

    Severity HIGH

    Risk Data breach — encrypted embedding store or audit log exfiltrated by an external attacker or insider.

    Rationale Embeddings are pseudonymous personal data and a breach would be reportable under UK GDPR Art 33. Likelihood is low because of layered controls (encryption at rest and in transit, role-scoped access, MFA); severity is high because a breach would erode resident trust and trigger ICO notification.

  • R4

    Likelihood MEDIUM

    Severity HIGH

    Risk Mission creep — the system is repurposed beyond the stated lawful basis (e.g. used for marketing analytics, footfall research, or employee monitoring).

    Rationale Without explicit contractual and technical limits, operators may be tempted to query the embedding store for purposes the residents were not notified of. Severity is high because new purposes would lack a lawful basis. Likelihood is medium without explicit controls.

  • R5

    Likelihood LOW

    Severity MEDIUM

    Risk Disproportionate retention — embeddings or events kept longer than necessary for the stated purpose.

    Rationale Our retention defaults (72hr ring buffer / 30d embeddings / 90d events) are aligned with ICO guidance and contractually fixed. Likelihood is low because retention is enforced by automated deletion jobs, not operator discipline.

  • R6

    Likelihood MEDIUM

    Severity MEDIUM

    Risk Lack of transparency — residents and visitors are not adequately informed that ReID processing takes place.

    Rationale Signage and privacy notices are the operator's responsibility. Severity is medium because a transparency failure undermines lawful basis but is remediable. Likelihood is medium without explicit operator training.

  • R7

    Likelihood MEDIUM

    Severity HIGH

    Risk Inaccurate match — a similarity hit links the wrong individual to an event, leading to operational decisions (e.g. exclusion) based on faulty evidence.

    Rationale No biometric system is perfect. Severity is high because operational consequences (challenge, exclusion) directly affect the data subject. Likelihood is medium and managed via the human-in-the-loop requirement.

  • R8

    Likelihood LOW

    Severity HIGH

    Risk Sub-processor breach — a downstream provider (compute, storage, monitoring) is compromised and exposes data we hold with them.

    Rationale Sub-processors are UK or EU only, contractually bound by a DPA, and listed publicly on /sovereignty#sub-processors. Likelihood is low because of vendor diligence; severity is high because a breach would still be reportable.

Step 6 — Measures to address the risks.

Each control is keyed to the risks it addresses. A single risk is typically held by multiple layered controls.

  • C1

    Addresses R1 · R3 · R8

    Control Encryption at rest (AES-256-GCM) and in transit (TLS 1.3)

    Detail All embedding stores, event logs and audit trails encrypted at rest with AES-256-GCM. All inter-service traffic encrypted in transit with TLS 1.3. No plaintext anywhere in the data path.

  • C2

    Addresses R1 · R2

    Control Embedding versioning and rotation

    Detail Model version recorded with every embedding. Embeddings written by an older model version are queryable but not minable for cross-version inversion. Rotation cadence: minor (calibration) every 6 months, major (new architecture) on demand.

  • C3

    Addresses R3 · R5

    Control Layered retention policy with automated deletion

    Detail Default retention: 72-hour on-camera ring buffer for raw frames; 30-day embedding window; 90-day event log; 12-month audit trail. Deletion is performed by automated jobs, not operator discipline. Custom retention shorter than default is contractually possible; longer requires fresh DPIA.

  • C4

    Addresses R3 · R4 · R8

    Control Role-scoped access controls + MFA

    Detail Operators see only their assigned cameras. Reviewers see only flagged events in their window. Admins see audit logs. Nobody sees raw embedding vectors via the UI. MFA enforced site-wide (TOTP plus WebAuthn). All access logged.

  • C5

    Addresses R3 · R4 · R7

    Control Immutable audit log

    Detail Every read of the embedding store, every match query, every operator decision, written to an append-only audit log retained 12 months. The customer's DPO can request the log at any time without notice.

  • C6

    Addresses R2 · R7

    Control Demographic bias monitoring

    Detail Model performance broken out by available demographic proxies (only those inferable from the operator's existing resident roster, never from the embedding itself). Per-group false-positive rates published to the customer's DPO each quarter.

  • C7

    Addresses R7

    Control Human-in-the-loop requirement on consequential actions

    Detail No automated exclusion or enforcement action permitted on a match alone. Every consequential action requires explicit operator confirmation and is logged with the operator's identity. Contractually enforced.

  • C8

    Addresses R4

    Control Purpose-limitation in contract and code

    Detail Customer contract limits use to the documented purpose. Technical limits (no marketing-API query path, no employee-monitoring path) prevent the most common mission-creep vectors. Repurposing requires fresh DPIA.

  • C9

    Addresses R6

    Control Transparency artefacts (signage, privacy notice template)

    Detail We provide a signage template and a privacy notice template the operator can drop into their existing comms. Both name the controller, the lawful basis and the retention period in plain English.

  • C10

    Addresses R8

    Control Sub-processor diligence and annual review

    Detail All sub-processors UK or EU. Each bound by a DPA flowing through our customer commitments. Annual review at contract anniversary; renewals not auto-pushed without review. List public at /sovereignty#sub-processors.

  • C11

    Addresses R5 · R6

    Control DSAR / right-to-erasure workflow

    Detail Subject access requests routed via the operator to us within 7 days; we resolve against embedding store and event log within 14 days, leaving the operator a 9-day buffer for the 30-day window. Erasure removes the embedding cluster and tombstones the event log entries.

  • C12

    Addresses R3 · R8

    Control Incident response procedure aligned to ICO 72-hour window

    Detail Customer notified within 24 hours of confirmed breach; ICO notification drafted by us, filed by the customer; 72-hour window met by default. Post-incident review feeds back into controls.

Step 7 — Sign-off and outcomes.

Residual risk. With the Step 6 controls in place, the residual risk profile is:

  • R1 (embedding reverse-engineering) — residual low.
  • R2 (model bias) — residual low / medium, monitored quarterly.
  • R3 (data breach) — residual low.
  • R4 (mission creep) — residual low given C8.
  • R5 (disproportionate retention) — residual low.
  • R6 (lack of transparency) — residual low if signage and notice are deployed; PLACEHOLDER — confirm.
  • R7 (inaccurate match) — residual low given the C7 human-in-the-loop requirement.
  • R8 (sub-processor breach) — residual low.

No residual risk is rated high; prior consultation with the ICO under UK GDPR Art 36 is not required.

Measures approved by

PLACEHOLDER — DPO name

Data Protection Officer

Signature: ____________________

Date: ____ / ____ / ________

Residual risks accepted by

PLACEHOLDER — SIRO name

Senior Information Risk Owner

Signature: ____________________

Date: ____ / ____ / ________

Consulted

PLACEHOLDER — names of others consulted (resident reps, joint controllers, ICO if applicable)

Consultees

Next review

PLACEHOLDER — annual by default; sooner on material change

Review cadence

Read the rest of the compliance pack.

This template sits alongside the DSPT v8 self-assessment and the sovereignty proof. Together the three are the evidence pack a BTR procurement team needs in front of them before signing the pilot.