Privacy, Fairness, Accountability and Transparency in Computer Vision

The rapid adoption of Generative AI has increased the need to protect Personally Identifiable Information, while still enabling organizations to extract value from their data. This abstract explores whether PII can be effectively anonymized for GenAI use cases, considering the requirements of GDPR and CCPA, including data minimization, purpose limitation, individual rights, and the distinction between anonymization and pseudonymization. It examines techniques such as synthetic data, while addressing the growing risk of re identification through AI models and data linkage. The key question is whether anonymized data can remain useful for GenAI while providing a level of protection that is technically robust and legally defensible.

We witness dramatic improvements in AI performance which fundamentally change the field of computer vision. But performance is not all: in order to support transparency and accountability in high-stakes applications, we need reliable explanations of their predictions. In this keynote, I present counterfactual explanations of AI models. Ranging from image classification to outlier detection and clustering methods, I illustrate how we can leverage the underlying model structure for reliable explanations.

• A Privacy Study of Sparse Collaborative Inference.    Maximilian Andreas Hoefler, Karsten Mueller, and Wojciech Samek

• Strategic Unlearnable Example Portfolios for Accountable Image Privacy.    Yizhou Chen and Meng Zhang

• HuC-VideoMAE: Human-Centric Video Masked Autoencoding from synthetic data.    Ricardo Pizarro, Roberto Valle, José M. Buenaposada, Luis M. Bergasa, and Luis Baumela

• Attributes Should Come from Images, Not Class Names: Distribution-Conditioned Attribute Selection for Vision-Language Models.    Gautam Rajendrakumar Gare, Jia Shi, Zhiqiu Lin, Deepak Pathak, John Galeotti, and Deva Ramanan

• Beyond Classification: Task-Dependent Learnability under Privacy-Motivated Image Transformations.    Leon Ranke, Wolfgang Hübner, Ronny Hug, Michael Arens, and Jürgen Beyerer

• State Space Models for Zero-Bootstrapping Homomorphic Image Classification.    Abhiram Srivatsa Kadaba

• Lensless Gaze Is Not Private by Default: Auditing Identity Leakage Across Disclosure Surfaces.    Rahul Vimalkanth, and Kaushik Mitra

• CIFA: Contextual-Intersectional Fairness Auditing for Hidden Subgroup Discovery in Face Analysis.    Nazia Aslam, Khalid Adnan Alsayed, Thomas B. Moeslund, and Kamal Nasrollahi

• SegWave: Wavelet-Driven Segmentation of Tampered Regions.    Siddhi Pravin Lipare, Vishesh Kumar, and Akshay Agarwal

• Controlled Auditing of Skin-Tone Shortcut Sensitivity in Human Action Recognition.    Ana Băltăreţu, Pascal Benschop, Justin Dauwels, and Jan van Gemert

• PRISM: Privacy-Aware Federated Periocular Identification under Extreme Label Skew for Mobile Devices.    Paula Delgado-Santo, Ruben Tolosana, and David Solans Noguero

• Mitigating Surrogate Overfitting for Membership Inference with Self-Supervised Models.    Sinan Mutlu, Savas Ozkan, and Mete Ozay

• Beyond Real vs Fake: Comparing GAN-Synthesized Faces and Diffusion-Based Face Reconstructions.    Ioana-Alexandra Tonu, Sebastian Tonu, and Otilia Zvorişteanu

• Soft Redaction of Image Provenance via Zero-Knowledge Proofs.    Muhammad Awan, and John Collomosse

• Multi-Level Knowledge Sharing in Federated Palmprint Authentication.    Jamal Seyedmohammadi, Pai Chet Ng, and Konstantinos N. Plataniotis

• Cen-FedRAL: Bridging Centralized Reinforcement Learning Training and Federated Active Learning for Dynamic Batch Acquisition on Resource-Constrained Devices.    Ioannis Lazaridis, Athanasios Psaltis, Anastasios Dimou, and Petros Daras

• Revisiting Attribution by Customization: A Study of Similarity-Based Attribution for Text-to-Image Diffusion Models.    Aveen Dayal, Jianbo Ma, Siqi Pan, Lovekesh Vig, and Andrea Fanelli

• Align Once to Explain: Feature Alignment for Scalable B-cosification of Foundational Vision Transformers.   Raphael Maser, Siddhartha Gairola, Sukrut Rao, and Bernt Schiele

• Culture in Action: Evaluating Text-to-Image Models through Social Activities.   Sina Malakouti, Boqing Gong, and Adriana Kovashka

• Debiasing Vision - Language Models without Catastrophic Forgetting.    Ryota Ishizaki, Yusuke Kuwana, and Go Irie

• Defending from GeoLocalization through Adversarial Road Trips.   Niccolò Niccoli, Federico Becattini and, Lorenzo Seidenari

• Layout-Agnostic Human Sensing with Illuminance Sensor Array   Hiroto Tsunoda, Hibiki Hariguchi, Yu Mitsuzumi, Akisato Kimura, Kiyoharu Aizawa, and Go Irie

• LoCo-Bot: Intrinsically Grounded Concept Bottleneck Models in a Single Forward Pass.    Sangwon Kim, Kyoungoh Lee, In-su Jang, and Kwang-Ju Kim

• Synchronized RGB-D Inpainting for Privacy-Aware 3D Scene Graph Construction.    Gawtam Chithra Ramesh, Püren Güler, Hiba Alqaysi, Marcus Valtonen Örnhag, Héctor Caltenco, Erdal Akin, and Kayode Sakariyah Adewole

Provenance records facts about digital content - who created it, how it was made, and what has happened to it. Open standards such as C2PA are increasingly used to carry these facts for content authenticity, but provenance can do much more than help us distinguish real from synthetic media. It can also provide infrastructure through which creators express how their work may be used, enabling better attribution, signalling creator preferences such as opt-out from AI training, and ultimately supporting licensing and value exchange when content is reused. In this talk I will explore this emerging role of provenance as infrastructure for digital rights. I will describe our work creating ZOETROPE - a provenance-based decentralised content exchange, and touch on recent advances using zero-knowledge proofs to selectively disclose provenance for privacy and personality-rights use cases. I will also discuss the challenge of ensuring that rights signals survive real-world distribution, including robust watermarking and signposting approaches for durable and interoperable provenance.

TBD