vaayaa Inclusive Education
Technical risk assessment

Computer Vision Systems Risk Assessment: Model Drift, Access, Privacy and Service Continuity

A worked example of how a team running a camera-based vision system finds, scores and controls risks to accuracy, image privacy, access, service continuity and supplier dependence.

Updated November 2026 8 hazards assessed Review every 6 months Highest risk: High to Medium
See the Computer Vision Systems course

Executive summary

This assessment shows how a team running a computer vision system can find, rate and control the risks that affect its accuracy, the privacy of the images it captures and its availability. Eight hazards were chosen because they are plausible in everyday work with cameras, models and cloud services. Model drift, unauthorized access to camera feeds and exposure of stored images carry the highest inherent scores, because a wrong detection or a leaked frame can cause real harm. Each hazard is scored before controls and again after existing and added controls. Most fall to Medium or Low once controls are in place, but model drift and access risks stay at Medium because the environment changes and attackers adapt. The first actions are a labeled drift test set, multi-factor sign in for all accounts, a retention limit on stored images and a documented rollback path for each model. The example is for teaching, so every score illustrates method rather than a real organization.

Focus area
Computer vision systems, image analysis and machine learning operations
Setting
Development and operation of camera-based inspection or monitoring systems for a mid-sized organization
Method
Five by five likelihood and impact risk register

Scope

This assessment covers an example computer vision pipeline, from camera capture and edge devices through model training, inference services, storage of images and results, and dashboards used by operators. It considers risks to accuracy, privacy, availability and trust. It does not cover any real deployment or legal ruling. Jurisdiction, camera locations, hardware vendors and whether people or screens appear in images are assumptions made for teaching, and they must be confirmed before any real use.

Context

The example system uses fixed cameras on a production line to flag defects. Edge devices run a trained model, and a central service stores images and results. A small machine learning team trains and updates the model, an operations team monitors alerts, and an IT group manages devices, networks and accounts. Some components come from outside suppliers, including a cloud platform and a pretrained model. Each risk is scored before controls to show its natural level, then again after existing controls and the additional controls proposed.

Objectives, assets and stakeholders

The objective is to keep the system accurate, protect the images and results it handles, keep it available to operators, and meet the privacy and security duties that apply. The assets in scope are camera feeds and stored images, trained models and their training data, edge devices, network links, cloud accounts, alert logs and the reputation of the team. Stakeholders include the machine learning and IT staff, the operators who act on alerts, the people who might appear in images, the owners of the production line, suppliers of hardware and cloud services, and any regulator with jurisdiction. Clear roles matter: every risk has one owner role, and the engineering director keeps the register current and reports to the steering group each quarter.

Risk matrix: before and after controls

Each cell shows hazard numbers placed by likelihood and impact. Compare the two grids to see how controls move each hazard.

Before controls

Inherent risk

Risk before controls: rows are likelihood from almost certain to rare, columns are impact from insignificant to catastrophic. Cells list hazard numbers.
1
Insignificant		 2
Minor		 3
Moderate		 4
Major		 5
Catastrophic
5 Almost certain
4 Likely Likelihood 4, impact 3, score 12, high: H04 Likelihood 4, impact 4, score 16, high: H01
3 Possible Likelihood 3, impact 2, score 6, medium: H07 Likelihood 3, impact 3, score 9, medium: H05 Likelihood 3, impact 4, score 12, high: H06 Likelihood 3, impact 5, score 15, high: H02 H03
2 Unlikely Likelihood 2, impact 5, score 10, high: H08
1 Rare

After controls

Residual risk

Risk after controls: rows are likelihood from almost certain to rare, columns are impact from insignificant to catastrophic. Cells list hazard numbers.
1
Insignificant		 2
Minor		 3
Moderate		 4
Major		 5
Catastrophic
5 Almost certain
4 Likely
3 Possible
2 Unlikely Likelihood 2, impact 2, score 4, low: H04 H05 H07 Likelihood 2, impact 3, score 6, medium: H01 H03 H06 H08 Likelihood 2, impact 4, score 8, medium: H02
1 Rare
  • Low 1 to 4
  • Medium 5 to 9
  • High 10 to 16
  • Critical 17 to 25

Hazard register

A summary first, then the detail for each hazard.

Summary of hazards with scores before and after controls
No.HazardBeforeAfterOwner
H01 Model drift and false detections in changing conditions 16 High 6 Medium Machine learning lead
H02 Unauthorized access to camera feeds and stored images 15 High 8 Medium IT manager
H03 Exposure of personal data in images and training sets 15 High 6 Medium Data protection lead
H04 Vulnerable edge devices and misconfiguration 12 High 4 Low Network engineer
H05 Service disruption from device, network or cloud outage 9 Medium 4 Low Operations manager
H06 Malware and phishing against workstations and model repositories 12 High 6 Medium Security lead
H07 Third-party model, hardware and cloud dependency 6 Medium 4 Low Procurement manager
H08 Insider misuse and gaps in logging and incident response 10 High 6 Medium Engineering director
H01

Model drift and false detections in changing conditions

Operational Before 16 After 6

Lighting, camera angle, product changes or dust reduce accuracy over time, so the model misses defects or raises false alarms without anyone noticing.

Who or what is affected: Operators, production staff and customers receiving the product

Causes

Changing light, wear on equipment, new product variants and training data that no longer match the line.

Consequences

Missed defects reach customers, false alarms slow the line and operators lose trust in alerts.

Existing controls

  • Operators can override an alert and note the reason.
  • Accuracy is checked by a sample review each month.

Control gaps

Sample reviews are small and happen monthly, so drift can build for weeks before it is seen.

Additional controls

  • Prevent: Fix camera mounts and lighting, and log any change that could affect the image.
  • Detect: Run a labeled drift test set on each model release and alert when accuracy drops.
Likelihood 4 x impact 4 before; 2 x 3 after Owner: Machine learning lead Target: Within 45 days Early warning: Drop in labeled test accuracy compared with the last approved model.
H02

Unauthorized access to camera feeds and stored images

Technology and data Before 15 After 8

Stolen passwords, shared accounts or misconfigured permissions let someone view live feeds or download stored images and results they should not see.

Who or what is affected: Staff who may appear in images, operators and the data protection lead

Causes

Reused passwords, long-lived access rights and permissions copied from earlier projects.

Consequences

Privacy breach, misuse of images to learn about products or people, and loss of trust.

Existing controls

  • Users sign in with individual accounts.
  • Image storage is restricted to the internal network.

Control gaps

Access rights are reviewed only when staff leave, so unused rights remain active.

Additional controls

  • Prevent: Require multi-factor sign in and role based access to every feed and storage location.
  • Detect: Log all access to images and review unusual downloads each week.
  • Respond: Disable any account that shows signs of misuse until the review is complete.
Likelihood 3 x impact 5 before; 2 x 4 after Owner: IT manager Target: Within 30 days Early warning: Access to camera feeds or stored images by accounts not on the approved list.
H03

Exposure of personal data in images and training sets

Legal and compliance Before 15 After 6

Images that show people, plates, screens or documents are kept longer than needed or copied into training sets without review, so personal information is exposed or used beyond its purpose.

Who or what is affected: People who appear in images, the team and any regulator with jurisdiction

Causes

No retention limit, copies made for experiments and unclear rules on what may be kept.

Consequences

Unlawful processing claims, notices to people affected, penalties where they apply and reputational harm.

Existing controls

  • Images are stored in a restricted location with a fixed folder structure.
  • Training data is collected for a stated purpose and recorded in a data log.

Control gaps

Retention is not set for backups or training copies, so old images persist beyond their purpose.

Additional controls

  • Prevent: Mask or blur people and screens at capture where the purpose allows it.
  • Recover: Delete copies from training sets and backups when a request or deadline requires it.
Likelihood 3 x impact 5 before; 2 x 3 after Owner: Data protection lead Target: Within 45 days Early warning: Stored images or training folders older than the agreed retention limit.
H04

Vulnerable edge devices and misconfiguration

Technology and data Before 12 After 4

Cameras and edge computers run old firmware, default settings or open ports, so attackers can take control of a device, alter its output or use it to reach the network.

Who or what is affected: Operators, IT staff and anyone relying on alerts from the line

Causes

Default settings, slow patch cycles and devices installed without a configuration record.

Consequences

Hijacked feeds, false or silent outputs, and a route into the wider network.

Existing controls

  • Devices sit on a separate network segment.
  • Firmware updates are installed when the vendor releases them.

Control gaps

Firmware is updated by hand, so some devices are months behind the current version.

Additional controls

  • Prevent: Change default passwords, close unused ports and disable services not needed for inspection.
  • Detect: Scan device configurations monthly and report any drift from the approved baseline.
Likelihood 4 x impact 3 before; 2 x 2 after Owner: Network engineer Target: Within 60 days Early warning: Edge devices running out of date firmware or a model version not on the approved list.
H05

Service disruption from device, network or cloud outage

Operational Before 9 After 4

A power loss, failed network link, full storage or cloud outage stops inspection or monitoring, so defects go unchecked or the line must stop while operators work by hand.

Who or what is affected: Production staff, customers and the operations team

Causes

Single points of failure, power and network faults and storage that fills without warning.

Consequences

Lost output, delayed orders, unchecked product and extra cost to recover.

Existing controls

  • Edge devices keep a local copy of the last approved model.
  • A cloud status page is checked at the start of each shift.

Control gaps

Manual inspection has not been rehearsed, and the restore test is not yet scheduled.

Additional controls

  • Respond: Switch to manual inspection using a written procedure when the service is down.
  • Recover: Test a full restore of the service and its configuration each quarter.
Likelihood 3 x impact 3 before; 2 x 2 after Owner: Operations manager Target: Within 60 days Early warning: Storage above the agreed fill level or repeated loss of connection from an edge device.
H06

Malware and phishing against workstations and model repositories

Technology and data Before 12 After 6

Phishing emails or infected files on engineers' workstations allow attackers to steal credentials, tamper with model code or plant a backdoor in a model that is then deployed to the line.

Who or what is affected: Machine learning staff, operators relying on the model and customers

Causes

Convincing phishing, unpatched software and downloads of unverified pretrained models.

Consequences

Compromised model behavior, data theft and costly rebuilds.

Existing controls

  • Email filtering blocks known malicious attachments.
  • Model code is kept in a version controlled repository with peer review.

Control gaps

Pretrained models from public sources are not scanned before use.

Additional controls

  • Prevent: Require signed model artifacts and a second review before any model is deployed.
  • Detect: Use endpoint monitoring and alert on unexpected changes to model files.
  • Respond: Isolate an infected workstation and revoke its credentials at once.
Likelihood 3 x impact 4 before; 2 x 3 after Owner: Security lead Target: Within 30 days Early warning: Unexpected change to a model file or repeated failed sign in on an engineer account.
H07

Third-party model, hardware and cloud dependency

Supplier Before 6 After 4

A supplier changes its pretrained model, camera firmware or cloud terms without notice, or fails to deliver support, so the system behaves differently or cannot be maintained.

Who or what is affected: Engineering team, procurement staff and operators

Causes

Vendor lock in, unclear license terms and reliance on one pretrained model.

Consequences

Unexpected changes in accuracy, delays and extra cost to replace parts or services.

Existing controls

  • Contracts include a notice period for major changes.
  • Key suppliers are listed in a supplier register.

Control gaps

Supplier changes are not tested before they reach the live line.

Additional controls

  • Prevent: Pin model and firmware versions and test any supplier update before rollout.
  • Respond: Keep a second camera supplier option and a documented exit plan for the cloud platform.
Likelihood 3 x impact 2 before; 2 x 2 after Owner: Procurement manager Target: Within 90 days Early warning: Supplier notice of a model, firmware or service change that has not yet been tested.
H08

Insider misuse and gaps in logging and incident response

People Before 10 After 6

A staff member copies images, changes labels or disables logs for personal gain or through carelessness, and no one notices because logs are incomplete or the response plan has not been rehearsed.

Who or what is affected: Owners, customers, colleagues and people in stored images

Causes

Broad rights, weak oversight and logs that are kept locally and not reviewed.

Consequences

Stolen or corrupted data, lost evidence and slow recovery from a breach.

Existing controls

  • Staff sign a confidentiality agreement on joining.
  • Access to training data is limited to named engineers.

Control gaps

Logs are kept on each device rather than centrally, and no one reviews them on a set schedule.

Additional controls

  • Detect: Centralize logs for image access, label edits and admin actions, with alerts for unusual activity.
  • Respond: Run a tabletop exercise of an insider incident each year and update the playbook afterward.
Likelihood 2 x impact 5 before; 2 x 3 after Owner: Engineering director Target: Within 45 days Early warning: Bulk download or label edit by a single account outside normal working hours.

Rating method and evaluation

Each risk is rated on two five-point scales. Likelihood runs from 1 Rare to 5 Almost certain, and impact runs from 1 Insignificant to 5 Severe. Multiplying the two gives a score from 1 to 25, which falls into one of four bands: Low from 1 to 4, Medium from 5 to 9, High from 10 to 16 and Critical from 17 to 25. The score is a guide for ordering work, not a measurement. Inherent risk is the score before any controls. Residual risk is the score after the existing and added controls. Impact is judged on harm to people whose images are held, lost output, cost to recover and damage to trust, and a privacy or security risk is never rated down only because it has not yet happened. Anything still High needs senior attention and a dated plan, anything Medium needs routine controls with a named owner, and anything Low is accepted and watched. The people who build and operate the system should challenge the ratings, and each rating is reviewed whenever a warning indicator reaches its threshold.

Likelihood, impact and rating bands

Likelihood

  • 5 Almost certain: Expected to happen in most circumstances.
  • 4 Likely: Will probably happen in many circumstances.
  • 3 Possible: Could happen at some time.
  • 2 Unlikely: Not expected, but it could happen.
  • 1 Rare: Only in exceptional circumstances.

Impact

  • 1 Insignificant: No lasting effect. Handled within routine work.
  • 2 Minor: Small effect that is quickly corrected.
  • 3 Moderate: Noticeable effect that needs managed effort to correct.
  • 4 Major: Serious effect on people, delivery or reputation.
  • 5 Catastrophic: Severe or lasting effect that threatens the whole activity.

Bands (likelihood x impact)

  • Low (1 to 4): Accept and monitor.
  • Medium (5 to 9): Manage with routine controls.
  • High (10 to 16): Senior attention and a treatment plan.
  • Critical (17 to 25): Act first. Do not proceed until reduced.

Monitoring, escalation and review

Review this assessment every 6 months, and sooner when an indicator below reaches its threshold.

Early-warning indicators and escalation thresholds
Indicator to watchEscalate when
Drop in labeled test accuracy compared with the last approved model.Escalate to the machine learning lead when accuracy falls by the agreed margin, and to the steering group if it stays low for two releases.
Access to camera feeds or stored images by accounts not on the approved list.Escalate to the security lead at once, and suspend the account until the review is complete.
Stored images older than the agreed retention limit.Escalate to the data protection lead when any are found, and to the operations manager if they are not removed within 14 days.
Edge devices running out of date firmware or a model version not on the approved list.Escalate to the IT manager within one working day, and isolate the device if it is reachable from outside the network.
Supplier status notices or outages affecting the cloud platform.Escalate to the engineering director at once, and switch to manual inspection if service is not restored within the agreed window.

Contingency arrangements

If a serious event occurs, such as a confirmed data exposure, a model that causes repeated false alarms or a long outage, the engineering director pauses the affected service and moves operators to manual inspection. The IT manager isolates compromised accounts and devices, and the model is rolled back to the last approved version. Images are preserved for investigation under the access limits, and the data protection lead decides on any notice to affected people or regulators, as required in the relevant jurisdiction. Supplier contacts are used for their own outages, and staff and operators are told what they need to know early and plainly.

Prioritized action plan

  1. Build a labeled drift test set and run it on every model release, owner machine learning lead, within 45 days.
  2. Enforce multi-factor sign in and role based access to camera feeds and stored images, owner IT manager, within 30 days.
  3. Set a retention limit for stored images and a regular deletion check, owner data protection lead, within 45 days.
  4. Keep a documented rollback path and a tested restore for every model and dataset, owner engineering director, within 60 days.
  5. Centralize logging with alerts for unusual access and adopt a written incident playbook, owner security lead, within 30 days.

Conclusion

Reliable computer vision depends on steady habits: labeled test sets, drift checks, controlled access, limits on stored images, tested rollback and logs that someone actually reads. The highest scores sit with drift, access and image exposure, and each has actions that lower the residual score. Treat the register as a living record, review it on the stated cycle, and seek qualified advice for any real privacy, safety or legal question.

Assumptions and limitations

Assumptions

  • Cameras capture industrial parts, and staff are not identified in images, but this must be confirmed in any real setting.
  • The team uses managed devices, a central account directory and a version-controlled model repository.
  • Jurisdiction is not specified, so privacy duties are described in general terms only.

Limitations

  • The scores are examples for learning and do not describe any real system or vendor.
  • Privacy, safety and sector rules for camera systems vary and change over time, so specialist review is needed.
  • This is educational content and does not replace a formal site-specific assessment or licensed advice.

Related frameworks to study

  • Threat modeling with a structured method such as STRIDE
  • Hierarchy of controls
  • Three lines model of risk ownership
  • Model cards and dataset documentation practice
  • Basic incident response lifecycle

Technical Risk Assessment

China, India, United States, Indonesia, Pakistan, Nigeria, Brazil, Bangladesh, Russia, Japan, Mexico, Philippines, Vietnam, Egypt, Turkey, Germany, United Kingdom, France, Italy, Canada, South Korea, Australia, Spain.

For Learning Reference Only: This assessment demonstrates how risk assessment principles are applied in a practical setting, providing a useful reference for learning and understanding. Actual assessments may be adapted to suit specific site conditions and circumstances.

Keep learning

Study the subject behind this assessment.

Computer Vision Systems course Read the best practices for this course

More Technical risk assessments

Hourly Active: 39,499 Learners