Speaker: Dr. Ajmal Mian
From: The University of Western Australia
Abstract
In this presentation, I will explore the manipulation of text-to-image (T2I) generative models through backdoors. I will present our Backdoor Attack on Generative Models (BAGM), which infuses the generated images with subtle manipulative details by injecting backdoors at three stages of the generative process: tokenizer, language model, or the image generator. For evaluation, we use a marketing scenario as the target domain and show that BAGM increases the bias towards target outputs more than fivefold without affecting untriggered model behavior. Beyond intentional manipulation, T2I models naturally contain bias which can propagate unfair social representations or push controversial agendas. We propose an evaluation methodology to quantify general biases in T2I generative models, without any preconceived notions. We assess four state-of-the-art T2I models and compare their baseline bias characteristics to their respective variants, where certain biases have been intentionally induced using BAGM. We propose three evaluation metrics: (i) Distribution bias, (ii) Jaccard hallucination and (iii) Generative miss-rate to measure biases under general and task-oriented conditions. Our method is available as a web application for measuring bias in any T2I model. In a follow up work, we expose the possibility of a dynamic and efficient exploitation of T2I model bias by targeting the language embeddings. By leveraging vector algebra, our technique enables convenient control over the severity of output manipulation and as a by-product, achieves a form of precise prompt engineering to generate images which are implausible with text prompts. Finally, we show a constructive application of this method for model debiasing and report compelling qualitative and quantitative results..
For more info, please follow this link.
Read More