From Privacy Chains to ChainShield: Structured Privacy Risks and Defense in Vision-Language Models

Recent advances in vision-language models have resulted in improved accuracy in visual question answering (VQA) tasks. However, their robustness remains limited when faced with out-of-distribution data containing unanswerable questions. In this study, we first construct a simple randomised VQA dataset, incorporating unanswerable questions from the VQA v2 dataset, to evaluate the robustness of a state-of-the-art VQA model. Our findings reveal that the model struggles to predict the “unknown” answer or provides inaccurate responses with high confidence scores for irrelevant questions. To address this issue without retraining the large backbone models, we propose Cross Modal Augmentation (CMA), a model-agnostic, test-time-only, multi-modal semantic augmentation technique. CMA generates multiple semantically-consistent but heterogeneous instances from the visual and textual inputs, which are then fed to the model, and the predictions are combined to achieve a more robust output. We demonstrate that implementing CMA enables the VQA model to provide more reliable answers in scenarios involving unanswerable questions, and show that the approach is generalisable across different categories of pre-trained vision language models.

With the wide application of machine learning technology, more and more sensitive data were used to develop the machine learning model. Unfortunately, machine learning systems have been shown to be vulnerable to various privacy leakage attacks, such as membership inference attacks and model inversion attacks. These privacy threats break down the confidentiality of the training data; what’s more, the leakage of sensitive privacy will reduce the enthusiasm of data owners for sharing data with the machine learning system. Without enough training data, more seriously, it will hinder the application of machine learning. Therefore, it requires analyzing strategies for privacy risks to help data owners pre-assess the candidate dataset and assist data owners in implementing reasonable privacy control. However, the systematic privacy risk assessment is still absent from the data owner’s perspective.This paper investigates and analyzes machine learning privacy risks to understand the relationship between training data properties and privacy leakage. Based on this recognition, we introduce a privacy risk assessment scheme based on the clustering distance of training data. Our clustering distance-based method can reflect the privacy risk level of a different individual data record. And then, we combine both existing privacy analysis based on data features and our clustering distance-based method to investigate privacy risks systematically. Our experiments showed that our clustering distance method and other set properties are tightly related to privacy leakage. And data owners can pre-assess the privacy risks of candidate datasets before uploading or sharing them to the machine learning model. If needed, decide to re-choose the dataset to reduce privacy risks to an acceptable level.

Visual question answering (VQA) research has garnered increasing attention in recent years. It is considered a visual Turing test because it requires a computer to respond to textual questions based on an image. Expertise in computer vision, natural language processing, knowledge understanding, and reasoning is required to solve the problem of VQA. Most techniques employed for VQA consist of models that are developed to learn the combination of image and question features along with the expected answer. The techniques chosen for image and question feature extraction and combining the features change with each model. This method of teaching a model of the question–answer pattern is ineffective for queries that involve counting and reasoning. This approach also requires considerable resources and large datasets for the training. The general VQA datasets feature a restricted number of items as responses to counting questions ([Formula: see text]), and the distribution of the answers is not uniform. To investigate these issues in VQA, we created synthetic datasets that could be modified to adjust the number of objects in the image and the amount of occlusion. Specifically, a zero-shot learning VQA system was devised for counting-related questions that provide answers by analyzing the output of an object detector and the query keywords. Using synthetic datasets, our model generated 100% correct results. Testing on the benchmark datasets task directed image understanding challenge (TDIUC) and TallyQA-simple indicated that the proposed model matched the performance of the learning-based baseline models. This methodology can be used efficiently for counting VQA questions confined to certain domains when the number of items to be counted is significant.

Visual Question Answering (VQA) has attracted much attention recently in both natural language processing and computer vision communities, as it offers insight into the relationships between two relevant sources of information. Tremendous advances are seen in the field of VQA due to the success of deep learning. Based upon advances and improvements, the Affective Visual Question Answering Network (AVQAN) enriches the understanding and analysis of VQA models by making use of the emotional information contained in the images to produce sensitive answers, while maintaining the same level of accuracy as ordinary VQA baseline models. It is a reasonably new task to integrate the emotional information contained in the images into VQA. However, it is challenging to separate questionguided-attention from mood-guided-attention due to the concatenation of the question words and the mood labels in AVQAN. Also, it is believed that this type of concatenation is harmful to the performance of the model. To mitigate such an effect, we propose the Double-Layer Affective Visual Question Answering Network (DAVQAN) that divides the task of generating emotional answers in VQA into two simpler subtasks: the generation of non-emotional responses and the production of mood labels, and two independent layers are utilized to tackle these subtasks. Comparative experimentation conducted on a preprocessed dataset to performance comparison shows that the overall performance of DAVQAN is 7.6% higher than AVQAN, demonstrating the effectiveness of the proposed model.We also introduce more advanced word embedding method and more fine-grained image feature extractor into AVQAN and DAVQAN to further improve their performance and obtain better results than their original models, which proves that VQA integrated with affective computing can improve the performance of the whole model by improving these two modules just like the general VQA.

We propose a model for free-form visual question answering (VQA) from human brain activity. The task of VQA is leading to an answer given an image and a question about the image. Given brain activity data measured by functional magnetic resonance imaging (fMRI) and a natural language question in terms of the viewed image, the proposed method can provide an accurate natural language answer with the VQA algorithm. Visual questions selectively approach various areas of an image such as objects and backgrounds. As a result, a more detailed understanding of the image and complex reasoning are typically needed than general image captioning models. In this paper, we propose a method of answering a given question about a viewed image from fMRI data based on the VQA algorithm. We estimate the relation between fMRI data and visual features extracted from viewed images. Based on the relationship, we convert fMRI data into visual features. Finally, the proposed method can answer to a visual question from fMRI data measured while subjects are viewing images. Experimental results show that the proposed method enables accurate answering for questions about viewed images.

The discussion about what can Artificial Intelligence (AI) contribute to games has been running for a long time, however, recent advances in AI show promise of providing new kind of experiences for players and new tools for game developers. In contrast with the traditional Finite State Machine for interaction and response, we consider the scenario of Visual Question & Answering (VQA) - the automatic answering of a textual question about an image. VQA is a tool that can enrich possible answers by combining both visual and textual information. It is also trivial to extrapolate to a game setting without going away from the training domain. In this Work In Progress, we present two original prototypes designed to explore the potential of VQA in games and discuss preliminary findings originated through a Wizard of Oz (WOz) pilot study using VQA to investigate how people interact with such an AI algorithm.

Visual question answering (VQA) is a prevalent task in real-world, and plays an essential role in helping the blind understand the physical world. However, due to the real-world complexity, VQA test samples may come from a different distribution from the training data, resulting in unavoidable performance degradation. This similar issue also exists in the image recognition field, in which one most recent effective solutions is a test-time adaptation (TTA). TTA adapts a trained model at test time using only test samples, which provides a new idea to alleviate the analogous issue in VQA. However, naively introducing existing TTA methods ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">e.g.,</i> test-time entropy minimisation) into VQA is imperfect and achieves only marginal performance gain. The reason is that prior methods do not consider the special nature of the VQA problem and ignore that 1) the biased samples in the dataset may have negative effects on test-time model adaptation, and 2) the model may have captured the biases in the dataset. In this paper, we propose Test-time Debiased Self-supervised (TDS) learning objectives for VQA model adaptation. Specifically, we minimise the entropy for those unbiased test samples. To identify these samples, we construct a negative sample for each test sample, and regard the test samples as unbiased if the output answers are different when feeding the test sample and the counterpart negative sample into the VQA model. Meanwhile, we also remove those samples with high prediction entropy from adaptation, making the test-time gradients more reliable. To hinder the model from excessively fitting the superficial correlations of the biased sample, we adopt the biased samples and the counterpart negative samples to assist the adaptation. Extensive experiments on the VQA-CP v1 and VQA-CP v2 datasets demonstrate the effectiveness of our TDS.

Visual question answering: Datasets, algorithms, and future challenges

Users face various privacy risks in smart homes, yet there are limited ways for them to learn about the details of such risks, such as the data practices of smart home devices and their data flow. In this paper, we present Privacy Plumber, a system that enables a user to inspect and explore the privacy "leaks" in their home using an augmented reality tool. Privacy Plumber allows the user to learn and understand the volume of data leaving the home and how that data may affect a user's privacy -- in the same physical context as the devices in question, because we visualize the privacy leaks with augmented reality. Privacy Plumber uses ARP spoofing to gather aggregate network traffic information and presents it through an overlay on top of the device in an smartphone app. The increased transparency aims to help the user make privacy decisions and mend potential privacy leaks, such as instruct Privacy Plumber on what devices to block, on what schedule (i.e., turn off Alexa when sleeping), etc. Our initial user study with six participants demonstrates participants' increased awareness of privacy leaks in smart devices, which further contributes to their privacy decisions (e.g., which devices to block).

Visual Question Answer (VQA), the natural language question of Visual images, has become popular in the field of artificial intelligence. At present, most of the VQA models extract the whole image features, which consume a large amount of computation and have a complex structure. In this paper, we propose a VQA method based on question attention model. Firstly, the Convolutional Neural Networks (CNN) is used to extract image features from the input images, and the question text is processed by the Long Short-Term Memory (LSTM). Then, we design a question attention module to let the learning algorithm focus on the most relevant features of the input text. According to question features, our method utilities the attention module to add the corresponding weights to the image features and extract the meaningful information for the generation of answer sequence words. Our method performed significantly better than the LSTMQ+I model on the MS COCO visual question answer (VQA) dataset with an accuracy improvement of 2%.

The task of visual question answering (VQA) is receiving increasing interest from researchers in both the computer vision and natural language processing fields. Tremendous advances have been seen in the field of computer vision due to the success of deep learning, in particular on low- and midlevel tasks, such as image segmentation or object recognition. These advances have fueled researchers' confidence for tackling more complex tasks that combine vision with language and high-level reasoning. VQA is a prime example of this trend. This article presents the ongoing work in the field and the current approaches to VQA based on deep learning. VQA constitutes a test for deep visual understanding and a benchmark for general artificial intelligence (AI). While the field of VQA has seen recent successes, it remains a largely unsolved task.

Understanding the diagnosis and management of external eye disease: Gurwood AS: Br J Optom Disp 1995;3:155–163

We study the problem of visual question answering (VQA) in images by exploiting supervised domain adaptation, where there is a large amount of labeled data in the source domain but only limited labeled data in the target domain, with the goal to train a good target model. A straightforward solution is to fine-tune a pre-trained source model by using those limited labeled target data, but it usually cannot work well due to the considerable difference between the data distributions of the source and target domains. Moreover, the availability of multiple modalities (i.e., images, questions and answers) in VQA poses further challenges in modeling the transferability between various modalities. In this paper, we address the above issues by proposing a novel supervised multi-modal domain adaptation method for VQA to learn joint feature embeddings across different domains and modalities. Specifically, we align the data distributions of the source and target domains by considering those modalities both jointly and separately. Extensive experiments on the benchmark VQA 2.0 and VizWiz datasets demonstrate that our proposed method outperforms the existing state-of-the-art baselines for open-ended VQA in this challenging domain adaptation setting.

A survey of methods, datasets and evaluation metrics for visual question answering

In this paper, we propose a novel large scale Visual Question Answering (VQA) dataset, which aims at real environment applications. Existing VQA datasets either require high constructing labor costs or have only limited power for evaluating complicated scene understanding ability involving in VQA tasks. Moreover, most VQA datasets do not tackle scenes containing object occlusion, which could be crucial for real-world applications. In this work, we propose a synthetic multi-view VQA dataset along with a dataset generation process. We build our dataset from three real object model datasets. Each scene is observed from multiple virtual cameras, which often requires a multi-view scene understanding. Our dataset requires relatively low labor cost and in the meantime, have highly complicated visual information. In addition, our dataset can be further adapted to users’ requirements by extending the dataset setup. We evaluated two previous multi-view VQA methods on our datasets. The results show that both 3D understanding and appearance understanding is crucial to achieving high performance in our dataset, and there is still room for future methods to improve. Our dataset provides a possible way for bridging the VQA methods aiming at CG dataset with real-world applications, such as robot picking tasks.