What is Model Drift?

Model Drift - Definition

Model drift refers to a change in the performance of a machine learning model over time caused by shifts in input data, feature distributions, labels, or the relationship between features and labels. The literature distinguishes between data drift (a change in the input distribution), concept drift (a change in the input-label relationship), and label shift (a change in the label distribution). See the survey by Gama et al. (2014, ACM Computing Surveys) and the taxonomy proposed by Webb et al. (2016, DMKD).

In image and video anonymization, model drift manifests as a degradation in the effectiveness of detecting objects that must be blurred - primarily faces and vehicle license plates. Changes in lighting, new license plate formats, masks, glasses, increased compression artifacts, or camera upgrades create a mismatch between training data and production data. This reduces recall and increases the risk of exposing personal data.

The Role of Model Drift in Video and Image Anonymization

In systems such as Gallio PRO, deep learning models detect faces and license plates and then apply blurring. Model drift can lead to missed detections or inaccurate object localization, resulting in missing blur or blur applied to an area that is too small. From a GDPR perspective, this poses a risk of violating the principle of integrity and confidentiality under Article 5(1)(f) and the requirements for data protection by design and by default under Article 25 of the GDPR.

Gallio PRO does not perform real-time anonymization and operates in batch mode. This makes drift control easier through periodic validation on reference datasets and model updates without the risk of immediate impact on live video streams.

Types and Symptoms of Model Drift in Anonymization

For data protection officers and technical teams, distinguishing the type of drift is crucial, as it determines the appropriate response. The table below provides a concise overview.

Key Parameters and Metrics for Drift Detection

Monitoring model drift requires objective measurements. In object detection, both quality metrics and distributional statistics are used. Below are the most common ones, along with sources and brief formulas.

• Recall and precision in object detection, computed on annotated reference datasets. The reporting standard is mAP calculated over IoU thresholds from 0.5 to 0.95 in steps of 0.05 (COCO; Lin et al., 2014).
• PSI - Population Stability Index, widely used to detect distribution drift. Formula: PSI = Σ_i (p_i − q_i) · ln(p_i / q_i), where p_i is the production share and q_i the reference share (Siddiqi, 2012).
• Classical distribution tests - the Kolmogorov-Smirnov test for continuous variables and the chi-square test for categorical variables, used to assess the statistical significance of distribution changes.
• KL divergence - a measure of divergence between distributions P and Q: D_KL(P‖Q) = Σ P(x) ln(P(x)/Q(x)) (Kullback-Leibler, 1951).
• Wasserstein distance - useful for comparing histograms of continuous image features and embeddings.
• FID - Fréchet Inception Distance, used to assess changes in the distribution of visual representations in embedding space (Heusel et al., 2017).

Monitoring Model Drift Without Collecting Personal Data

Gallio PRO does not collect logs containing face or license plate detections and does not process logs with personal data. Nevertheless, model drift can be monitored using non-personal and aggregated data.

• Synthetic and control datasets - periodic evaluation on annotated test image sets designed to contain no identifiable individuals. A good benchmark is validation procedures aligned with ISO/IEC 23894:2023, which include AI risk assessment.
• Aggregated telemetry without PII - distributions of bounding box sizes, average confidence scores, and histograms of sharpness and lighting after anonymization. The data are aggregated and non-identifying.
• Feature distribution tests - PSI, KL, and KS computed on technical image features (e.g., luminance, noise) and embeddings that cannot be used to reconstruct identities.

Drift Mitigation Strategies and Corrective Actions

An effective drift management plan should combine MLOps processes with regulatory requirements. NIST AI RMF 1.0 (2023) as well as ISO/IEC 22989:2022 and ISO/IEC 23894:2023 provide frameworks for the AI lifecycle and risk management.

• Periodic retraining - updating face and license plate detection models using new data, with full validation of mAP and recall on control datasets.
• Domain-specific augmentation - simulating lighting conditions, motion blur, compression artifacts, vignetting, and new license plate formats to reduce sensitivity to data drift.
• Threshold calibration - adjusting confidence thresholds and blur mask margins to minimize missed detections. All changes must be documented and tested on reference datasets.
• Model and data versioning - maintaining repositories of models, configurations, and test datasets, with full audit trails of changes and metric results in line with ISO/IEC 23894:2023.

Challenges and Compliance

Variability in recording quality, heterogeneous camera setups, and evolving local regulations increase the risk of insufficient blurring. In its Guidelines 3/2019 (version 2.0 of 29 January 2020), the EDPB notes that identifiable elements such as a person’s image (face) or vehicle license plates may constitute personal data if they enable identification of a natural person. Their anonymization must therefore be effective and robust. The GDPR requires appropriate technical and organizational measures and risk assessment (Articles 5 and 25).

In parts of Western Europe, blurring license plates is required in certain contexts by law or established practice. In Poland, interpretations in case law differ, alongside guidance from the national DPA (UODO) and the EDPB emphasizing the protection of identifiability. In this regulatory environment, managing model drift is an essential element of due diligence and compliance documentation.

Technologies Supporting Drift Detection and Reduction

Face and license plate detection relies on supervised convolutional neural networks and transformers. Maintaining these systems requires tooling for data distribution analysis and continuous validation. The table below summarizes selected best practices.

Normative References and Sources

The following high-authority sources underpin the definitions and recommendations cited in this text.

• GDPR - Regulation (EU) 2016/679, Articles 5 and 25; Recital 26.
• EDPB, Guidelines 3/2019 on the processing of personal data through video devices, version 2.0 of 29 January 2020.
• ISO/IEC 22989:2022 - Artificial intelligence - Concepts and terminology.
• ISO/IEC 23894:2023 - Artificial intelligence - Risk management.
• NIST AI Risk Management Framework 1.0, 2023.
• Gama, J. et al., A Survey on Concept Drift Adaptation, ACM Computing Surveys, 2014.
• Webb, G.I. et al., Characterizing Concept Drift, Data Mining and Knowledge Discovery, 2016.
• Lin, T.-Y. et al., Microsoft COCO: Common Objects in Context, 2014 - mAP/IoU definitions.
• Siddiqi, N., Credit Risk Scorecards, 2012 - PSI definition.
• Kullback, S., Leibler, R.A., On Information and Sufficiency, 1951 - D_KL definition.
• Heusel, M. et al., GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium, 2017 - FID.