While Multimodal Large Language Models (MLLMs) offer strong perception and reasoning capabilities for image-text input, Visual Question Answering (VQA) focusing on small image details still remains a challenge. Although visual cropping techniques seem promising, recent approaches have several limitations: the need for task-specific fine-tuning, low efficiency due to uninformed exhaustive search, or incompatibility with efficient attention implementations. We address these shortcomings by proposing a training-free visual cropping method, dubbed FOCUS, that leverages MLLM-internal representations to guide the search for the most relevant image region. This is accomplished in four steps: first, we identify the target object(s) in the VQA prompt; second, we compute an object relevance map using the key-value (KV) cache; third, we propose and rank relevant image regions based on the map; and finally, we perform the fine-grained VQA task using the top-ranked region. As a result of this informed search strategy, FOCUS achieves strong performance across four fine-grained VQA datasets and three types of MLLMs. It outperforms three popular visual cropping methods in both accuracy and efficiency, and matches the best-performing baseline, ZoomEye, while requiring 3 – 6.5 × less compute.

Federated learning (FL) when deployed in an Internet of Things (IoT) ecosystem can facilitate the collaborative training of a global model involving different IoT local systems. However, there are a number of challenges in such deployments, and examples include single point of failure / attack, lack of fault tolerance, vulnerability to collusion attacks and accuracy loss. Therefore, we propose a privacy-preserving serverless FL scheme for IoT based on secure multi-party computation. Specifically, in our scheme, no central sever is required to coordinate the generation of global models. In doing so, we avoid the single point of failure / attack limitation. We also mitigate the fault tolerance limitation by using secret sharing. Finally, we provide a formal security proof that demonstrates the resilience of our scheme against collusion attacks, thereby establishing its effectiveness in achieving robust data privacy. Simulations are also implemented to show that our scheme does not suffer from accuracy loss.

Federated learning (FL) allows multiple users to collaboratively train global machine learning models by keeping their data sets local. However, the existing privacy-preserving FL schemes suffer from several limitations, e.g., loss of accuracy, high communication/computation cost, failure to support dynamic users, and insecurity against collusion attacks. To solve these limitations, we propose a lightweight privacy-preserving FL scheme based on a dual-server architecture. Our scheme involves only lightweight cryptographic operations, i.e., hash and symmetric encryption operations, and it has low communication overhead. Thus, it is computationally lightweight and round-efficient. Further, it allows users to join/quit an FL task and it is accuracy-lossless. We formally prove that our scheme remains secure even in case of collusion attacks. In particular, if an attacker colludes with one of the servers and all the users who participate in an FL task except two, the privacy of user gradients stays unviolated. The reported experimental results demonstrate that our scheme incurs only a marginal increase in total communication overhead compared to the FL scheme without any privacy protection. In terms of computation overhead, the cost per user remains stable as the number of users grows, while the cost for the server is comparable to that of the FL scheme without any privacy protection.