Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Minimizing Risk Through Minimizing Model-Data Interaction: A Protocol For Relying on Proxy Tasks When Designing Child Sexual Abuse Imagery Detection Models

Thamiris Coelho, Leo S. F. Ribeiro, João Macedo, Jefersson A. dos Santos, Sandra Avila

TL;DR

CSAI detection is hampered by restricted access to CSAI data. The authors formalize Proxy Tasks as related sub-tasks paired with CSAI-specific data constraints and present a data-centric protocol for designing such tasks with LEA input. They validate the protocol via a case study on few-shot indoor scene classification, showing that a PMF-based embedding approach pre-trained on public data can achieve meaningful CSAI-related performance without training on CSAI weights. The results support the feasibility of Proxy Tasks to minimize model-data interaction and enable safer data triage for LEAs. The study also discusses ethical, practical, and future directions for building a library of Proxy Task models.

Abstract

The distribution of child sexual abuse imagery (CSAI) is an ever-growing concern of our modern world; children who suffered from this heinous crime are revictimized, and the growing amount of illegal imagery distributed overwhelms law enforcement agents (LEAs) with the manual labor of categorization. To ease this burden researchers have explored methods for automating data triage and detection of CSAI, but the sensitive nature of the data imposes restricted access and minimal interaction between real data and learning algorithms, avoiding leaks at all costs. In observing how these restrictions have shaped the literature we formalize a definition of "Proxy Tasks", i.e., the substitute tasks used for training models for CSAI without making use of CSA data. Under this new terminology we review current literature and present a protocol for making conscious use of Proxy Tasks together with consistent input from LEAs to design better automation in this field. Finally, we apply this protocol to study -- for the first time -- the task of Few-shot Indoor Scene Classification on CSAI, showing a final model that achieves promising results on a real-world CSAI dataset whilst having no weights actually trained on sensitive data.

Minimizing Risk Through Minimizing Model-Data Interaction: A Protocol For Relying on Proxy Tasks When Designing Child Sexual Abuse Imagery Detection Models

TL;DR

CSAI detection is hampered by restricted access to CSAI data. The authors formalize Proxy Tasks as related sub-tasks paired with CSAI-specific data constraints and present a data-centric protocol for designing such tasks with LEA input. They validate the protocol via a case study on few-shot indoor scene classification, showing that a PMF-based embedding approach pre-trained on public data can achieve meaningful CSAI-related performance without training on CSAI weights. The results support the feasibility of Proxy Tasks to minimize model-data interaction and enable safer data triage for LEAs. The study also discusses ethical, practical, and future directions for building a library of Proxy Task models.

Abstract

The distribution of child sexual abuse imagery (CSAI) is an ever-growing concern of our modern world; children who suffered from this heinous crime are revictimized, and the growing amount of illegal imagery distributed overwhelms law enforcement agents (LEAs) with the manual labor of categorization. To ease this burden researchers have explored methods for automating data triage and detection of CSAI, but the sensitive nature of the data imposes restricted access and minimal interaction between real data and learning algorithms, avoiding leaks at all costs. In observing how these restrictions have shaped the literature we formalize a definition of "Proxy Tasks", i.e., the substitute tasks used for training models for CSAI without making use of CSA data. Under this new terminology we review current literature and present a protocol for making conscious use of Proxy Tasks together with consistent input from LEAs to design better automation in this field. Finally, we apply this protocol to study -- for the first time -- the task of Few-shot Indoor Scene Classification on CSAI, showing a final model that achieves promising results on a real-world CSAI dataset whilst having no weights actually trained on sensitive data.
Paper Structure (19 sections, 7 figures, 2 tables)

This paper contains 19 sections, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Proxy Tasks and the Literature on CSAI Recognition
  3. Protocol for Developing CSAI Detection Solutions
  4. Case Study: Few-shot Learning for Indoor Scene Classification
  5. The Literature as it Relates to the Proxy Task
  6. Few-shot Learning Problem Definition
  7. Comparative Analysis
  8. Methodology
  9. Pre-training Datasets
  10. Fine-tuning Dataset
  11. Training and Evaluation Details
  12. Results
  13. Evaluation on Child Sexual Abuse Imagery
  14. Case Study Main Takeaways and Future Work
  15. Conclusion
  16. ...and 4 more sections

Figures (7)

  • Figure 1: The sensitive nature of CSAI requires that most research be done under a related substitute task before being applied to real data. Proxy Tasks combine a related sub-task of CSAI recognition with restrictions common in CSAI data; the figure shows examples of both aspects.
  • Figure 2: Diagram depicting the cycle of steps on our protocol. We believe that Proxy Tasks and CSAI evaluation datasets should be designed with a Data-Centric mindset; Proxy tasks are then used in lieu of real CSAI data to minimize model-CSAI interactions; real CSAI is used only through our partnerships with LEAs to evaluate final models.
  • Figure 3: Confusion matrix with the results of the model on the Places8 test set. The confusion matrix presents the accuracy (%) of the predicted classes following the FSL evaluation protocol. Ground truth labels are on the vertical axis; predicted labels on the horizontal axis.
  • Figure 4: t-SNE on two dimensions for the Places8 test set. Many children's room samples (blue dots) are closer to the bedroom prototype (orange star) than to the child's room (blue star). We observe the same phenomenon between classroom samples (pink dots) and child's room samples (blue dots).
  • Figure 5: Confusion matrix with the model's results on the CSAI samples classified into indoor. Values reported are the accuracy (%) for prediction per class. Ground truth labels are on the vertical axis; predicted labels on the horizontal axis.
  • ...and 2 more figures