TL;DR
This paper introduces a novel system that uses Generative AI (GenAI) to proactively search the internet for sensitive images that have already leaked, such as ID cards or driver's licenses. By combining image-based search with AI-driven filtering and OCR, the system effectively identifies and flags these leaks, offering remediation advice to mitigate risks.
What Problem Does This Paper Address?
In today's digital world, vast amounts of data are shared across various platforms, creating significant privacy and security concerns. While traditional Data Loss Prevention (DLP) systems focus on preventing data from leaving an organization's network, they are largely ineffective at detecting sensitive images that have already been exposed online. This gap is particularly problematic because images containing personal identifiable information (PII) can be shared without explicit textual labels, making them difficult to find with standard search methods. The paper aims to bridge this gap by developing a system that can proactively identify these leaked sensitive images on the internet.
Methods
The proposed system integrates six key modules:
- Input Module: Allows users to specify the type of sensitive data (e.g., driver's license, ID card) and country of origin they wish to search for.
- Generation Module (GenAI): Uses GenAI to create synthetic sensitive images based on standardized document layouts and user-defined prompts. This allows for the generation of realistic examples of sensitive documents without using real individuals' data.
- Search and Filter Module: Employs image-based search (using Selenium) to find candidate images online. It then uses BRISK feature descriptors with a Brute Force (BF) matcher, enhanced by Lowe's ratio test and a dynamically generated region mask, to filter out irrelevant matches and improve precision.
- Recognition Module: Utilizes a deep learning-based Optical Character Recognition (OCR) tool (PaddleOCR) to extract text and numbers from the filtered images. This recognized content is then verified against predefined sensitive data patterns using regular expressions.
- Detection and Marking Module: Categorizes detected sensitive images as "High Risk" or "Potential Risk" based on the completeness and format of the identified data.
- Report Generation Module: Consolidates all processed data, including flagged images, identified sensitive data, and risk levels, to generate reports. This module uses a local LLM (LLAMA 3.2) to provide actionable remediation recommendations for cyber security and privacy compliance.
Datasets
- Real-world web image retrieval: The system was evaluated using 610 real-world web images.
- Taiwan driver's license dataset: 80 publicly accessible images of Taiwan driver's licenses were used for a case study.
- IDNet dataset: A synthetic dataset containing identity documents from various countries was used to evaluate the Recognition Module. Specifically, a subset from three countries with publicly available checksum validation mechanisms was selected, totaling 17,937 images.
Key Insights and Ideas
- GenAI for synthetic data generation: Using GenAI to create realistic synthetic sensitive images is crucial for training and testing a proactive detection system, especially when real leaked data is scarce or sensitive to use.
- Layered approach for robust detection: Combining image-based search (to find images without explicit text) with OCR (to extract and verify text within images) provides a comprehensive strategy for identifying sensitive data leaks.
- Dynamic region masking for improved matching: Applying a dynamically generated mask based on the GenAI-generated reference layout significantly reduces false positives by focusing feature matching on structurally relevant regions.
- LLMs for actionable remediation: Integrating LLMs into the reporting module provides users with practical, context-aware recommendations for addressing detected data leaks, moving beyond simple detection.
Key Results
- High Precision and Recall: The system achieved a precision of 100% and a recall of approximately 93.1% in real-world evaluations.
- Accurate Recognition: The Recognition Module demonstrated strong performance on the IDNet dataset, achieving 100% precision and 99.7% recall across various identity document types.
- Effective Filtering: The dynamically generated mask reduced false matches by 70.4% in inter-class matching scenarios.
- Fidelity of GenAI outputs: The system's GenAI-generated images showed significantly lower pHash distances compared to other widely used text-to-image models, indicating superior structural and perceptual fidelity.
- Real-world leak discovery: The system successfully identified actual data leaks, including instances of severe access control vulnerabilities allowing unauthorized access to driver's license data.
Limitations and Future Work
- Computational Intensity of Generation: The Generation Module, while producing high-fidelity outputs, is computationally intensive. Future work could explore lighter-weight generative approaches.
- Latency in Report Generation: Local LLM inference for report generation introduces latency. Future iterations could incorporate retrieval-augmented generation (RAG) for faster responses.
- Robustness to Image Transformations: The current feature matching (BRISK) might be less effective under significant image rotations or scale changes. Future work will include robustness testing with synthetic transformations.
- Human Factors: The authors acknowledge the importance of human factors, suggesting ongoing security awareness training and clear procedural guidelines as complementary measures to automated detection.
- Broader Evaluation: Future work could involve more diverse evaluations, particularly with low-quality or noisy inputs, to further enhance system robustness and generalizability.
Summary updated