Reliable behavior control is central to deploying Large Language Models (LLMs) on the web. Activation steering offers a tuning-free route to align attributes (e.g., truthfulness) that ensure trustworthy generation. Prevailing approaches rely on coarse heuristics and lack a principled account of where to steer and how strongly to intervene. To this end, we propose Position-wise Injection with eXact Estimated Levels (PIXEL), a position-wise activation steering framework that, in contrast to prior work, learns a property-aligned subspace from dual views (tail-averaged and end-token) and selects intervention strength via a constrained geometric objective with a closed-form solution, thereby adapting to token-level sensitivity without global hyperparameter tuning. PIXEL further performs sample-level orthogonal residual calibration to refine the global attribute direction and employs a lightweight position-scanning routine to identify receptive injection sites. We additionally provide representation-level guarantees for the minimal-intervention rule, supporting reliable alignment. Across diverse models and evaluation paradigms, PIXEL consistently improves attribute alignment while preserving model general capabilities, offering a practical and principled method for LLMs' controllable generation. Our code is available at https://anonymous.4open.science/r/PIXEL-Adaptive-Steering-95DC

• OOD data significantly impacts model performance, resulting in decreased prediction accuracy and reliability. • The continual tabular contrastive learning (C-TCL) method effectively manages OOD data while maintaining computational efficiency. • C-TCL achieves superior performance on CPU hardware, making it more accessible than GPU-based alternatives. • Experimental results from eight diverse datasets demonstrate C-TCL’s effectiveness, particularly in classification tasks. • The matrix augmentation technique using full data representation improves model efficiency compared to traditional slice-based methods. • Simplified contrastive loss calculations with dot product reduces computational overhead while maintaining performance.

Out-of-distribution (OOD) prediction remains a significant challenge in machine learning, particularly for tabular data where traditional methods often fail to generalize beyond their training distribution.This paper introduces Tabular Continual Contrastive Learning (TCCL), a novel framework designed to address OOD challenges in tabular data processing.TCCL integrates contrastive learning principles with continual learning mechanisms, featuring a three-component architecture: an Encoder for data transformation, a Decoder for representation learning, and a Learner Head.We evaluate TCCL against 14 baseline models, including state-of-the-art deep learning approaches and gradient-boosted decision trees (GBDT), across eight diverse tabular datasets.Our experimental results demonstrate that TCCL consistently outperforms existing methods in both classification and regression tasks on OOD data, with particular strength in handling distribution shifts.These findings suggest that TCCL represents a significant advancement in handling OOD scenarios for tabular data.

The proliferation of deepfake technologies poses urgent challenges and serious risks to digital integrity, particularly within critical sectors such as forensics, journalism, and the legal system. While existing detection systems have made significant progress in classification accuracy, they typically function as black-box models, offering limited transparency and minimal support for human reasoning. This hinders their usability in real-world decision-making contexts, especially for non-expert users. We present DF-P2E (Deepfake: Prediction to Explanation), a novel multimodal framework that integrates visual, semantic, and narrative layers of explanation to make deepfake detection interpretable and accessible. The framework consists of three modular components: (1) a deepfake classifier with Grad-CAM-based saliency visualisation, (2) a visual captioning module that generates natural language summaries of manipulated regions, and (3) a narrative refinement module that uses a fine-tuned LLM to produce context-aware, user-sensitive explanations. We instantiate and evaluate the framework on the DF40 benchmark, the most diverse deepfake dataset to date. Experiments demonstrate that our system achieves competitive detection performance while providing high-quality explanations aligned with Grad-CAM activations. By unifying prediction and explanation in a coherent, human-aligned pipeline, this work offers a scalable approach to interpretable deepfake detection, advancing the broader vision of trustworthy and transparent AI systems for media forensics.

Large Language Models (LLMs) have revolutionised natural language processing tasks, particularly as chat agents. However, their applicability to threat detection problems remains unclear. This paper examines the feasibility of employing LLMs as a Network Intrusion Detection System (NIDS), despite their high computational requirements, primarily for the sake of explain-ability. Furthermore, considerable resources have been invested in developing LLMs, and they may offer utility for NIDS. Current state-of-the-art NIDS rely on artificial benchmarking datasets, resulting in skewed performance when applied to real-world networking environments. Therefore, we compare the GPT-4 and LLama3 models against traditional architectures and transformer-based models to assess their ability to detect malicious NetFlows without depending on artificially skewed datasets, but solely on their vast pre-trained acquired knowledge. Our results reveal that, although LLMs struggle with precise attack detection, they hold significant potential for a path towards explainable NIDS. Our preliminary exploration shows that LLMs are unfit for the detection of Malicious NetFlows. Most promisingly, however, these exhibit significant potential as complementary agents in NIDS, particularly in providing explanations and aiding in threat response when integrated with Retrieval Augmented Generation (RAG) and function calling capabilities.

Biometric authentication systems have become an inseparable part of society. This popularity is owing to the fact that biometric traits are immutable. However, applications using these sensitive biometric traits must treat this crucial information carefully. Otherwise, their leakage can threaten the security and privacy of an individual. This paper proposes a privacy-preserving biometric authentication system based on iris data. In the proposed framework, the homomorphic properties are exploited to process data while it is encrypted. There is no leakage of sensitive data throughout the entire process, even when utilizing the services of third-party cloud service providers (CSPs). Authentication is carried out fully within the encrypted domain. Experiments have been conducted to validate the robustness of the proposed framework. Additionally, the time complexity of the proposed framework is minimized compared to other state-of-the-art approaches.

In recent years, the rapid improvement of deep learning technologies, particularly Generative Adversarial Net-works, has led to the proliferation of high-quality synthetic facial images and videos, commonly known as deepfakes. This study aims to evaluate and compare the performance of three prominent deep learning models - ResN et, EfficientNet, and Xception - in detecting synthetic faces. Using the Deepfake Detection Challenge and FaceForensics++ datasets, we system-atically assess each model's capability to handle diverse and challenging scenarios, including blurred and dark images. Data augmentation techniques, such as random blurring, brightness adjustment, and contrast enhancement, were employed to im-prove the models' robustness. Additionally, we applied model- specific optimizations, including the integration of Squeeze-and- Excitation blocks in Res Net, compound scaling in EfficientNet, and multi-scale feature fusion in Xception. These enhancements significantly improved the models' accuracy and resilience against low-quality synthetic data. Our results indicate that EfficientNet and Xception outperform ResNet in both general and adverse conditions, with EfficientNet excelling in high-resolution image processing and Xception showing superior performance in fine- grained feature extraction. Furthermore, the introduction of pre- trained weights, multitask learning frameworks, and dynamic learning rate adjustments during training contributed to the models' enhanced performance.

Most of the devastating cyber-attacks are caused by insiders with access privileges inside an organization. The main reason of insider attacks being more effective is that they don't have many security barriers before they get into the critical resources of the system. Different machine learning techniques have been previously utilized to identify insider threats within cy-bersecurity domain whereas research done in predicting insider attacks is not significant. Moreover, machine learning models used for prediction and detection face a critical limitation as they require training on labeled datasets, rendering them less effective for real-time data streams which lack threat presence indicators. This work presents an unsupervised machine learning approach that predicts insider threat using behavior analysis for real-time threat data. Patterns are identified in user behavior, to make predictions about benign and malicious insiders. Features are selected by analyzing activities performed. Selected features are utilized to feed machine learning model which extracts anomalous behavior among users, using anomalies in their activity patterns followed by learning methods for threat detection. A dataset that contains selected features from CERT r4.2 is used to make predictions. The performance of Isolation Forest (iForest) is compared with other algorithms of the same category including One-class SVM, Local Outlier Factor (LOF) and DBSCAN to evaluate the new approach. The iForest shows the best performance accuracy 80 percent and recall 84.2 percent.

Cloud computing has become a significant part of people's daily lives over the last decade due to its cost-effective services, such as Infrastructure as a Service (IaaS), Platform as a Service (PaaS), and Software as a Service (SaaS). As a result, the cloud environment contains a massive amount of sensitive and confidential data, thus becoming a crucial target for attackers. Additionally, cloud forensics over the cloud environment has become complex and difficult due to the cloud's several challenges. These challenges include collecting volatile data, dependency on cloud service providers (CSPs), secure logs storage, and more. This paper proposes a Cloud-Based Data Volatility and CSP Reliance Eradication Framework (CDCEF) to mitigate data volatility and dependency on CSP in cloud forensics in IaaS. We also performed experiments on the Amazon Web Service (AWS) Lightsail - a widely used cloud platform by people globally, based on hypothetical cybercrime scenarios to support our framework. The significant benefits of this framework for investigators are that it allows them to collect volatile data without losing the integrity of the evidence and eradicates CSP reliance. Additionally, the framework provides another benefit of minimizing the usage of forensic tools.

Federated Learning (FL) allows for training machine learning models on decentralized data. However, FL has been prone to adversarial attacks. This paper examines the vulnerability of FL towards white-box attacks using the CIFAR10 dataset. In the study, we have used ResNet20 and DenseNet. In the study, the required perturbation is added to find the adversarial samples to fool the model. This decentralized approach to training can make it more difficult for attackers to access the training data, but it can also introduce new vulnerabilities that attackers can exploit. We conducted three types of white box attacks, i.e., Fast Gradient Sign Method (FGSM), Carlini-Wagner (CW), and DeepFool, and studied the model's behavior. We have presented the results of the model behavior considering the different scenarios.

Deep learning models have been widely employed in a multitude of security-critical contexts due to their impressive performance across diverse tasks. In this paper, we examine a specific form of assault on the Federated Learning (FL) framework. The utilization of Federated Learning (FL) has gained significant traction as a distributed training approach due to its ability to facilitate the processing of extensive datasets without the need for data sharing among users. After the training process of the model has been completed using data stored on local devices, solely the modified model parameters are transmitted to the central server. The approach of FL is characterized by its distribution across several components or entities. Hence, it is plausible for an individual to initiate an assault with the intention of manipulating the behavior of the model. This work presents a study on the implementation of a Backdoor attack, wherein a small number of malicious instances were introduced to assess the behavior of the model throughout the testing phase. The poisoning could be applicable to either a singular class or numerous classes. A series of experiments were undertaken using the widely used CIFAR10 dataset in order to modify the behavior of the model. It was determined that the anticipated performance of the model might be undermined with minimal impact on training accuracy.

The unlimited growth in the amount of multimedia content has shifted the global infrastructure to the cloud-based multimedia hosting. However, the high probability of security breaches of the content is increasing demand for secure solutions toward this end. One such feasible solution is to encrypt the content to unreadable form before outsourcing to the cloud-based servers. In this paper, the media information, specifically the video content is distributed into multiple random shares based on the permutation ordered binary number system. The information remains fully concealed without any leakage at the cloud servers. Even if the attacks endanger the integrity of the information, the proposed scheme is enriched with the capability of chalking out accurately the altered pixels. These tampered regions are subsequently reflected in the reconstructed video frames obtained at authentic entity end possessing the secret keys required to build back the original content. Moreover, the proposed scheme is capable of detecting temporal attacks on the video frames by employing sequence-based authentication bits. The robustness of the proposed scheme has been validated under different attack scenarios and the scheme is found to be performing satisfactorily well.

Message Locked Encryption (MLE) provides a way to achieve deduplication of data over the cloud in a secured and efficient manner. Various cryptographic protocols that are variants of the MLE scheme have been introduced, but they are either vulnerable to poison attack or consume a large amount of bandwidth. We introduce a new client-side deduplication protocol, called DICE (Dual Integrity Convergent Encryption), in which we perform tag checking and send the hash instead of the entire message over the network, resulting in a reduction of communication and computation cost without any loss of security. Our strategy is secure against both the erasure and the duplicate-faking attacks (known as a poison attack when performed together). Comparative analysis with other existing strategies validate the efficacy of the proposed protocol.

We often send our confidential information such as passport, credit card, social security numbers over email without concern about the security of email services. Existing network security mechanisms provide adequate security from external malicious adversaries and eavesdroppers, but they don't guarantee that the email service providers (ESPs) wouldn't or can't access our email data themselves, which in some cases could be highly confidential. One of the ways to protect email data from ESPs is to use Pretty Good Privacy (PGP) that has many limitations including key storage problem and dependability on third party services, making it cumbersome to use in practice. In this paper, we present S3Email method that provides email security against ESPs. The proposed method uses a cryptographic secret sharing technique in a novel way and encrypts the email metadata, body and attachments before the email is sent. In the proposed solution, the email sender and receiver must have at least two email accounts on the existing ESPs, which is not unusual today. Experiments and analysis show that the S3Email method provides information theoretic security with minimal computational overhead.

Developmental Disorders are chronic disabilities that have a severe impact on the day to day functioning of a large section of the human population. Recognizing developmental disorders from facial images is an important but a relatively unexplored challenge in the field of computer vision. This paper proposes a novel framework to detect developmental disorders from facial images. A spectrum of disorders constituting of Autism Spectrum Disorder, Cerebral Palsy, Fetal Alcohol Syndrome, Down syndrome, Intellectual disability and Progeria have been considered for recognition. The framework relies on Deep Convolutional Neural Networks (DCNN) for feature extraction. A new data-set comprising of images of subjects with these disabilities was built for testing the performance of the frame work. This model has been tested on different age groups, individual disabilities and has also been compared to a similar model that uses human intelligence to identify different developmental disorders. The results indicate that the model performs better than average human intelligence in terms of differentiating amongst different disabilities and is able to recognize subjects with these developmental disorders with an accuracy of 98.80%.

The benefits of high-end computation infrastructure facilities provided by cloud-based multimedia systems are attracting people all around the globe. However, such cloud-based systems possess security issues as third party servers become involved in them. Rendering data in an unreadable form so that no information is revealed to the cloud data centers will serve as the best solution to these security issues. One such image encryption scheme based on a Permutation Ordered Binary Number System has been proposed in this work. It distributes the image information in totally random shares, which can be stored at the cloud data centers. Further, the proposed scheme authenticates the shares at the pixel level. If any tampering is done at the cloud servers, the scheme can accurately identify the altered pixels via authentication bits and localizes the tampered area. The tampered portion is also reflected back in the reconstructed image that is obtained at the authentic user end. The experimental results validate the efficacy of the proposed scheme against various kinds of possible attacks, tested with a variety of images. The tamper detection accuracy has been computed on a pixel basis and found to be satisfactorily high for most of the tampering scenarios.

Security breaches are increasing day by day as the technology is innovating and opening new doors. There are various counter measures to keep a check on them and one such method is the watermarking. Region based watermarking is the approach where an important area of the original image is selected to hide the secret information. This is done to make the watermarking more robust to the various attacks and retain the commercial value of the image as any damage to this region would result in deterioration of the value of the image. One such quad tree based approach to select the region of interest (ROI) and then to utilize the properties of the singular value decomposition (SVD) transform to hide the watermark is being proposed here. The robustness of the methodology against the various attacks is validated by the experimental results.