Sunder Ali Khowaja

Federated Learning (FL) paradigm has been very popular in the implementation of 5G and beyond communication systems as it provides necessary security for the users in terms of data. However, the FL paradigm is still vulnerable to model inversion attacks, which allow malicious attackers to reconstruct data by using the trained model gradients. Such attacks can be carried out using generative adversarial networks (GANs), generative models, or by backtracking the model gradients. A zero-trust mechanism involves securing access and interactions with model gradients under the principle of “never trust, always verify.” This proactive approach ensures that sensitive information, such as model gradients, is kept private, making it difficult for adversaries to infer the private details of the users. This paper proposes a zero-trust based Block Encryption LAyer (BELA) module that provides defense against the model inversion attacks in FL settings. The BELA module mimics the Batch normalization (BN) layer in the deep neural network architecture that considers the random sequence. The sequence and the parameters are private to each client, which helps in providing defense against the model inversion attacks. We also provide extensive theoretical analysis to show that the proposed module is integratable in a variety of deep neural network architectures. Our experimental analysis on four publicly available datasets and various network architectures show that the BELA module can increase the mean square error (MSE) up to 194% when a reconstruction attempt is performed by an adversary using existing state-of-the-art methods

Consumer-based optical imaging systems are characterized as big data processing systems, which are drastically affected by atmospheric turbulences that add geometric distortions and blur effect to the images when used in outdoor condition. Physics-grounded simulators have been proposed recently to generate synthetic data but the generalization to the real-world turbulent images is not so good. In this paper, we combine the characteristics of image registration, deep neurofuzzy methods, and channel-attention based discriminative learning strategy to propose image registration, neurofuzzy based denoising, and deblurring network (RND2Net). The RND2Net is designed on a principle that it does not require turbulent image pairs (ground truth images) to train the network, which closely resembles the real-world situation used as consumer devices. The registration module focuses on the region-based fusion techniques while the denoising and deblurring module incorporates deep neurofuzzy network along with dense residual blocks and channel attention mechanism to train the network. The RND2Net is also designed to reduce the noise and blur effect from images, while generalizing on the down-stream tasks, such as text recognition. Experimental results show that the RND2Net yields better performance quantitatively as qualitatively on synthetic and real-world datasets in comparison to existing state-of-the-art methods

Satellite high aerial platform terrestrial integrated networks have become the hot topic these years, which have been regarded as the major part of the intelligence of things for future networks. During this work, we investigate the secrecy performance for rate-splitting multiple access-assisted satellite high aerial platform terrestrial integrated networks. Besides, imperfect hardware and channel estimation errors are considered in the secrecy networks. Moreover, to enhance the energy utilization efficiency, rate-splitting multiple access scheme is utilized into the considered network, which is prior to that of the non-orthogonal multiple access scheme. What’s more, to enhance the satellite transmission, multiple high aerial platforms are used to forward the transmission along with multiple eavesdroppers. In addition, the direct transmission link is not considered in the secrecy networks due to the heavy fading and obstacles. Relied on the former considerations, the exact and asymptotic analysis for the secrecy performance is further obtained to confirm the rightness of the analysis. Finally, some representative Monte Carlo simulations are carried out to validate the obtained results.

Self-supervised learning in the federated learning paradigm has been gaining a lot of interest both in industry and research due to the collaborative learning capability on unlabeled yet isolated data. However, self-supervised based federated learning strategies suffer from performance degradation due to label scarcity and diverse data distributions, i.e., data heterogeneity. In this paper, we propose the SelfFed framework for medical images to overcome data heterogeneity and label scarcity issues. The first phase of the SelfFed framework helps to overcome the data heterogeneity issue by leveraging the pre-training paradigm that performs augmentative modeling using Swin Transformer-based encoder in a decentralized manner. The label scarcity issue is addressed by fine-tuning paradigm that introduces a contrastive network and a novel aggregation strategy. We perform our experimental analysis on publicly available medical imaging datasets to show that SelfFed performs better when compared to existing baselines and works. Our method achieves a maximum improvement of 8.8% and 4.1% on Retina and COVID-FL datasets on non-IID datasets. Further, our proposed method outperforms existing baselines even when trained on a few (10%) labeled instances.

Post COVID-19 and recent advancement in terms of language models, researchers have shown a lot of interest in analyzing social media posts for analyzing mental state of the users. Social media platforms are the epitome of sharing individual thoughts and feelings through textual posts and linguistic cues. Therefore, the textual modality from social media posts can be leveraged for detecting early signs of stress, depression or other mental health conditions, accordingly. Existing methods mainly focus on the feature engineering, shallow learning, and employing of deep learning architectures to improve the mental state recognition performance. Seldom the study uses an established knowledge-base that is available to model mentalization and emotional aspect to improving the depression and stress recognition. In this regard, we propose emotion aware contrastive networks (EAC-net) that leverages the existing knowledge-base and propose some new ones to model the emotional and mentalization aspect in order to improve the recognition of stress and depression state from textual posts. Furthermore, we propose a feature-level fusion and weighting mechanism using GRUs and self-attention layers to weight and select the important features. Lastly, the EAC-Net uses a supervised contrastive learning strategy to train the network. The proposed method is evaluated on four publicly available datasets. Experimental results reveal that the EAC-Net achieves state-of-the-art results by outperforming baselines and existing methods by atleast 1.86%, 0.72%, 3.43%, and 3.64% on four publicly available datasets using F1-measure as the evaluation metric.

This paper reviews the Challenge on Blind Enhancement of Compressed Image at NTIRE 2024, which aims at enhancing the quality of JPEG images which are compressed with unknown quality factor. The challenge requires that the total size of codes and pre-trained model(s) cannot exceed 300 MB, since we encourage solutions for blind enhancement with generalized models, instead of separately training several models for each quality factor. In this report, we summarize the detailed settings of the challenge, the final results, and the solutions proposed by the participants. The challenge has 129 registered participants and received 13 valid submissions. Several teams (including all TOP 3 teams) have publicly released the codes (see Sec. 4). They gauge the state-of-the-art of blind quality enhancement of compressed image.

In past, due to data and model security concerns, modern communication systems mainly focus on the use of edge computing devices for enabling immersive applications and services. Federated learning is one of the preferred solutions but it stresses the computation capability of the edge devices for immersive applications. Much research is now focusing on split learning as an alternative due to its ability of performing joint training with limited computing resources. However, split learning is also vulnerable to data reconstruction, feature space hijacking, and model inversion attacks, which are quite common concerning immersive applications such as Metaverse. In this regard, we propose a ZEro-Trust Attack (ZETA) framework for data reconstruction and model inversion attacks for autonomous vehicles opting for split learning strategies. We propose the joint training of client, server, and shadow models for both the reconstruction and main task to fool existing methods. Our experimental results demonstrate that the proposed method is capable of reconstructing client’s data with an error of 0.0032. This study is proposed as a basis to design more sophisticated defense mechanisms for autonomous vehicles to protect user services in 5G/6G networks.

The fifth generation (5G) networks are designed to support a large range of diverse services with strict performance requirements. Studies suggest that, 5G uses machine learning technologies for variety of tasks ranging from network management, and resource optimization to automated services. The successful integration of 5G with machine learning has also led to the basis for 6G networks. However, the use of machine learning makes the 5G networks susceptible to adversarial attacks. A few works study the effect of differential privacy and adversarial attacks in the 5G systems let alone to provide the proposal of effective defense mechanism. This study proposes Denoising and Adversarial attack-based STacked AutoeNcoder (DASTAN) convolutional neural networks (CNN) to provide defense against a specific differential privacy attack, i.e. membership inference, optimized to detect the device or data distribution potentially used in the training process. DASTAN initiates an intentional attack to camouflage the characteristics of an authorized user from an adversary and uses a de noising stacked autoencoder to recover the information at service provider’s end for RF fingerprinting. The aim of RF fingerprinting is to validate the authenticity and identity of the device to preserve the privacy of wireless network. Experimental results demonstrate the efficacy of DASTAN-CNN, which reduces the attack success rate by up to 52.69% in comparison to the case where no defense strategy is employed. The DASTAN-CNN also achieves 75.29% authorized user recognition rate for RF fingerprinting while reducing the attack success rate to 39.23%, which shows the effectiveness in terms of trade-off efficiency.

This paper reviews the NTIRE 2023 challenge on image super-resolution (×4), focusing on the proposed solutions and results. The task of image super-resolution (SR) is to generate a high-resolution (HR) output from a corresponding low-resolution (LR) input by leveraging prior information from paired LR-HR images. The aim of the challenge is to obtain a network design/solution capable to produce high-quality results with the best performance (e.g., PSNR). We want to explore how high performance we can achieve regardless of computational cost (e.g., model size and FLOPs) and data. The track of the challenge was to measure the restored HR images with the ground truth HR images on DIV2K testing dataset. The ranking of the teams is determined directly by the PSNR value. The challenge has attracted 192 registered participants, where 15 teams made valid submissions. They achieve state-of-the-art performance in single image super-resolution.

This paper reviews the NTIRE 2023 challenge on image denoising (σ = 50) with a focus on the proposed solutions and results. The aim is to obtain a network design capable to produce high-quality results with the best performance measured by PSNR for image denoising. Independent additive white Gaussian noise (AWGN) is assumed and the noise level is 50. The challenge had 225 registered participants, and 16 teams made valid submissions. They gauge the state-of-the-art for image denoising.

The applications concerning vehicular networks benefit from the vision of beyond 5G and 6G technologies such as ultra-dense network topologies, low latency, and high data rates. Vehicular networks have always faced data privacy preservation concerns, which lead to the advent of distributed learning techniques such as federated learning. Although federated learning has solved data privacy preservation issues to some extent, the technique is quite vulnerable to model inversion and model poisoning attacks. We assume that the design of defense mechanism and attacks are two sides of the same coin. Designing a method to reduce vulnerability requires the attack to be effective and challenging with real-world implications. In this work, we propose simulated poisoning and inversion network (SPIN) that leverages the optimization approach for reconstructing data from a differential model trained by a vehicular node and intercepted when transmitted to roadside unit (RSU). We then train a generative adversarial network (GAN) to improve the generation of data with each passing round and global update from the RSU, accordingly. Evaluation results show the qualitative and quantitative effectiveness of the proposed approach. The attack initiated by SPIN can reduce up to 22% accuracy on publicly available datasets while just using a single attacker. We assume that revealing the simulation of such attacks would help us find its defense mechanism in an effective manner.

This paper reviews the NTIRE 2022 Challenge on Super-Resolution and Quality Enhancement of Compressed Video. In this challenge, we proposed the LDV 2.0 dataset, which includes the LDV dataset (240 videos) and 95 additional videos. This challenge includes three tracks. Track 1 aims at enhancing the videos compressed by HEVC at a fixed QP. Track 2 and Track 3 target both the super-resolution and quality enhancement of HEVC compressed video. They require x2 and x4 super-resolution, respectively. The three tracks totally attract more than 600 registrations. In the test phase, 8 teams, 8 teams and 12 teams submitted the final results to Tracks 1, 2 and 3, respectively. The proposed methods and solutions gauge the state-of-the-art of super-resolution and quality enhancement of compressed video. The proposed LDV 2.0 dataset is available at https://github.com/RenYang-home/LDV_dataset. The homepage of this challenge (including open-sourced codes) is at https://github.com/RenYang-home/NTIRE22_VEnh_SR.

Face recognition system is broadly used for human identification because of its capacity to measure the facial points and recognize the identity in an unobtrusive way. The application of face recognition systems can be applied to surveillance at home, workplaces, and campuses, accordingly. The problem with existing face recognition systems is that they either rely on the facial key points and landmarks or the face embeddings from FaceNet for the recognition process. In this paper, we propose a hierarchical network (HN) framework which uses pre-trained architecture for recognizing faces followed by the validation from face embeddings using FaceNet. We also designed a real-time face recognition security door lock system connected with raspberry pi as an implication of the proposed method. The evaluation of the proposed work has been conducted on the dataset collected from 12 students from Faculty of Engineering and Technology, University of Sindh. The experimental results show that the proposed method achieves better results over existing works. We also carried out a comparison on random faces acquired from the Internet to perform face recognition and results shows that the proposed HN framework is resilient to the randomly acquired faces.

The concept of connected things by involving emotional aspects has been raised as a new research issue which is known as "emotional IoT". The deeper interaction between object and human shows an importance to develop a system with either cognitive or affective capabilities such as emotion. While the existing works on real time emotion recognition mostly rely on facial data, there are a few works dealing with real time emotion recognition based on physiological data using pervasive devices. In this work, we propose a framework to recognize emotion based on human physiological signals using the pervasive wearable device. This framework opposed most of the works which employed sensors which are expensive and complex in arrangement. The challenge on using pervasive devices is the low accuracy due to the low sampling rate. The approach is implemented in an end-to-end soft real time emotion recognition system using smartphone and smartwatch devices. The performance of our system was evaluated under a common environment and proved the system applicability throughout everyday life.

User localization in wireless sensor networks have been given a great attention in recent times and also considered to be one of the promising applications. Many approaches for the same have been proposed based on range based and range free mechanisms for localizing the user. Similarly, tracking the user in a sensor field has also been given equal importance in association with localization. Techniques employing both these methods are mostly based on static anchor nodes or scheduling system which compromises on the lifetime of wireless sensor network. Considering the constraint of network lifetime this paper proposes localizing and tracking method with an activation scheme for tracking the mobile node efficiently while increasing the network lifetime. The system has been tested on two scenarios suggesting that the proposed method can provide the flexibility to the system which could be adjusted with reference to the user requirements. The first scenario suggests that all the nodes are activated only when the mobile user enters but as the user is localized all the nodes will get deactivated except the concerned nodes. The second scenario suggests that a predefined deployment strategy is provided with only 10% of activated nodes. The experimental results show that the proposed system achieves a better trade-off in terms of accuracy and computational complexity for single mobile node tracking.

With the convergence of smart technologies and advancement in electronic equipment the concept of smart home system swiftly escalates. The idea is to automate the home appliances according to the user requirements without human intervention. After a long tiring day and heavy workloads user will not be in a state of taking out its mobile phone and pressing the buttons for controlling home appliances. Several methods have been proposed in the design of such systems using sensors, biometrics and face detection. This paper proposes a method for detecting human emotions by taking into account the complete facial analysis, suggesting that the emotions can accurately be determined by analyzing eyes, nose and lips separately hence covering a wide range of emotions. The classification is carried out by acquiring the image of user followed by the face detection and segmenting the region of interests (ROI) i.e. eyes, nose and lips for further analysis of emotions. Principle Component Analysis (PCA) along with feature extraction techniques and Support Vector Machines (SVMs) are used for classification of emotion for the said automation system. Policies have been implemented in Java to simulate the home automation environment for testing and validation. At the instant this system has been tested on a single user with 4 basic emotions i.e. sad, anger, happiness and neutral, but this study can be a basis to develop an automated system with variety of emotions for multiple users.