How AITHENA advances anonymisation for CCAM applications?

Advancing Data Anonymization for CCAM Applications: Striking a Balance Between Privacy and Usability

Introduction

As Cooperative, Connected, and Automated Mobility (CCAM) technologies evolve, handling sensitive data securely while maintaining its usability for AI applications is critical. Data anonymization—removing identifiable information—is key to ensuring privacy while enabling robust AI development. However, traditional methods such as blurring introduce artifacts that compromise data quality, particularly in computer vision tasks. To address this, Vicomtech, as part of the AITHENA project, has developed cutting-edge anonymization approaches tailored for CCAM applications.

Challenges of Anonymization in CCAM

Anonymization typically modifies data to obscure identifiable features such as faces and license plates. However, such modifications often degrade data utility for training and validating AI models. This is particularly evident in blurring techniques, which, while computationally lightweight, introduce artifacts and remove significant image details. Generative AI presents a promising alternative, minimizing these issues but demanding higher computational resources.

Innovative Approaches to Anonymization

Vicomtech’s work in the AITHENA project has introduced two pioneering strategies for anonymizing data in CCAM applications, each with distinct strengths and trade-offs. The first approach relies on traditional blurring techniques, which are well-established in the field of data anonymization. This method obscures identifiable information such as faces and license plates by applying blur filters. It is computationally lightweight, making it suitable for scenarios where processing resources are limited, or rapid anonymization is essential. However, blurring removes significant details from the image and often introduces artifacts, such as pixelation or unnatural distortions. These issues can degrade the quality of the data, posing challenges for applications like computer vision training and validation, where high-quality inputs are critical. Additionally, a key limitation of blurring is its insufficient protection against re-identification. Even when the face is blurred, individuals can often be identified through other characteristics, such as their pose, body shape, clothing, or accessories, especially in scenarios where these elements remain consistent and recognizable.

The second and more innovative approach leverages the power of generative artificial intelligence (AI) to create anonymized images that retain much of their original utility. Generative AI models, enable the transformation of an original image into a visually realistic anonymized version without introducing the artifacts associated with blurring. This approach minimizes the loss of information by preserving the overall scene, including critical elements like the presence and posture of individuals. Such precision allows the data to remain highly effective for AI model development and testing.

To achieve this level of realism, Vicomtech employs advanced generative strategies. One method involves the use of image-to-image (Img2Img) models, which translate an image directly into its anonymized counterpart in a single step. Another strategy involves a sophisticated pipeline combining deep neural networks (DNNs) with specialized tasks, such as detecting people in the image, estimating masks to identify regions to anonymize, inpainting to fill gaps after anonymization, and generating synthetic representations of individuals. This approach ensures that the image retains its natural look and context while anonymizing sensitive information.

While generative techniques are highly effective, they require significant computational resources. Recognizing this challenge, Vicomtech is actively working on optimizing these processes to make them more scalable and practical for real-world applications. By focusing on these innovative methods, Vicomtech aims to strike a balance between data privacy and usability, ensuring that anonymized datasets remain valuable for advancing CCAM technologies.

You can watch the AITHENA Proven Anonymiser video on the Vicomtech YouTube channel: