As artificial intelligence continues to drive advancements in computer vision, particularly in areas such as image analysis, object detection, and facial recognition, the ability to accurately recognize patterns in visual data has become a central focus of research. However, alongside these advances, concerns about the privacy risks associated with the training data used in AI models have also gained prominence. Deep learning models, frequently employed in computer vision tasks, can unintentionally expose sensitive information from the data they are trained on, raising the need for comprehensive research into privacy-preserving techniques. This paper explores the intersection of AI-driven pattern recognition and the privacy risks involved in training models on image data. Existing studies show that attackers can exploit the gradients from deep learning processes to reconstruct original image data, including personal and identifiable information, such as facial features. By iteratively adjusting input data, attackers can minimize the difference between the gradients of the random and stolen data, leading to the full reconstruction of private images. Current privacy protection methods fall short of explaining the relationship between an attacker’s capacity to recover visual data and the structure of the targeted model. This paper introduces a novel privacy auditing framework that directly assesses the extent to which gradient-based attacks can reconstruct sensitive data. Unlike traditional methods, which mainly focus on mitigating privacy risks through model regularization or data obfuscation, our approach provides a systematic and quantitative evaluation of gradient leakage, filling a critical gap in existing privacy protection techniques. This paper investigates the relationships among reconstructed data, model gradients, and the original input data in the context of computer vision. By formalizing the connection between gradient similarity and data similarity, we propose a novel methodology that quantifies the vulnerability of deep learning models to data reconstruction attacks. Building on these insights, we propose a novel privacy auditing method aimed at evaluating the privacy risks associated with deep learning models used in pattern recognition for image data.
Digital transformation has become a new driving force for corporate development, widely embraced by managers. However, in the actual digital transformation process, some companies’ performances are characterized by “more talk, less action”. To measure the phenomenon, this paper innovatively constructs the Digital Disclosure Anomaly (DDA) indicator using Chinese A-share listed companies’ data from 2013 to 2021 and examines the impact of DDA on corporate innovation efficiency by Saying is Believing theory. We find that DDAs show a long-term negative impact on corporate innovation efficiency. Our mechanism analysis suggests that DDAs reduce innovation efficiency through resource misallocation, which is driven by two mechanisms: the manager squandering effect and the inefficient investment effect. Our results suggest that DDAs are harmful for corporate innovation efficiency by the Saying is Believing (SIB) effect and corporate DDA behavior is detrimental to sustainability in digital economy development.