Deepfake detection has advanced rapidly alongside progress in generative adversarial networks, diffusion models, neural rendering, and multimodal synthesis. However, strong performance on curated benchmark datasets does not necessarily translate into reliable performance in real-world environments, where manipulated media are affected by platform compression, reposting, post-processing, multimodal inconsistency, temporal drift, and evolving generation pipelines. This review examines deepfake detection from the perspective of in-the-wild robustness, multimodal reasoning, and cross-domain generalization. It first clarifies the conceptual foundations of in-the-wild and generalizable deepfake detection, then reviews the data ecosystem, including benchmark datasets, real-world media conditions, multimodal data gaps, annotation quality, and benchmark aging. The review further analyzes the major causes of detector failure under realistic deployment conditions, including generator shift, platform-induced degradation, modality mismatch, language and speaker variation, and broader contextual diversity. It then compares major methodological families, including artifact-based methods, deep representation-based approaches, transfer-based learning, domain adaptation, domain generalization, and multimodal detection strategies, with emphasis on their relevance to real-world robustness. Finally, the review discusses evaluation practices, external validation, realistic benchmarking, and future directions such as continually updated benchmarks, missing-modality-robust systems, uncertainty-aware detection, provenance integration, and foundation-model-based approaches. Overall, the review argues that reliable deepfake detection requires a shift from benchmark-centered optimization toward robust, multimodal, and deployment-oriented evaluation and design.
The fields of information retrieval and recommender systems have long evolved as distinct disciplines, each with its own methodologies and theoretical foundations. This survey provides a perspective on the convergence of these fields, a transformation catalyzed by the advent of large language models. We maintain that this convergence is far from a trivial overlap of tools, but it embodies a substantial conceptual and architectural unification. Our central contribution is a novel three-part taxonomy that explains the mechanisms of this merger leveraging large language models as (1) an embedded engine, (2) a semantic re-ranker, and (3) a response generator. We show how this trajectory culminates in a paradigm inversion, from weak to strong inversion, where the information-seeking task itself is reshaped. The unification is structurally consolidated through joint search and recommendation architectures and contextually curated by retrieval-augmented generation frameworks. This hypothesis is substantiated by a rigorous, quantitative analysis of scholarly publication trends, which provides empirical evidence of this rapid evolution. Furthermore, our analysis reveals a convergence operating at different velocities dictated by domain-specific latency constraints, while acknowledging critical operational vulnerabilities regarding hallucination and bias. A companion GitHub repository is publicly available and offers the complete data collection methodology, dynamic trend analyses, and a comprehensive, categorized list of the surveyed papers, which will be actively maintained as a living resource for the research community. We conclude that the distinction between retrieval and recommendation is becoming a historical artifact and call for the formal recognition of a new, integrated science of information discovery.
Multimodal Emotion Causality (MEC) studies why emotions arise in conversations by linking emotion-bearing turns to textual, acoustic, and visual evidence. As the field shifts from pair extraction to free-text rationales, evaluation practices have become fragmented across output interfaces. Link-level outputs expose alignment objects that can be matched to benchmark targets, whereas generative explanations are often assessed with non-comparable automatic metrics and subjective human or LLM judgments. Existing surveys emphasize tasks, datasets, and model families, but rarely connect extraction-style and generation-style outputs through their evaluation interfaces. An interface-centric synthesis frames MEC research as a transition from pair extraction toward verifiable grounded explanations. A capability matrix organizes datasets by linking supervision signals to the evidence units available for evaluation, while a bridgeability map distinguishes pointer-free narratives from outputs with machine-parseable evidence pointers. A Bridging Protocol is further presented as a minimal reporting protocol for mapping pointer-bearing explanations back to benchmark targets, with an accompanying open-source toolkit for reproducible parse-back and metric reporting. The survey concludes with open challenges in fine-grained temporal grounding, long-context causal attribution, low-resource adaptation, and reproducible evaluation.
The rapid progression of automotive technology through In-Vehicle Networks (IVNs) such as Controller Area Networks (CAN), Local Interconnect Network (LIN), FlexRay, Media Oriented Systems Transport (MOST), and Automotive Ethernet has significantly reshaped internal vehicle communication. This expansion in connectivity has led to new cybersecurity risks that directly affect passenger safety and the reliability of transportation infrastructure. These challenges highlight the need for resilient infrastructure for sustainable industrialisation. Many existing reviews in the IVN cybersecurity domain focus individually on specific protocols or attack categories and corresponding defense strategies rather than considering the entire framework. Thus, we conduct a Systematic Literature Review (SLR) on existing literature from 2010 to 2025 and examine cybersecurity developments related to IVNs. A structured taxonomy that integrates IVN protocols, major attack techniques, and datasets, along with cryptographic, conventional and Artificial Intelligence (AI)-based cybersecurity mechanisms, is presented. Our study adheres to PRISMA guidelines and provides a comprehensive and systematic analysis of the IVN cybersecurity domain. This provides a consolidated data synthesis of existing literature, offering deeper insights into current trends, research gaps, and emerging research directions in IVN cybersecurity. The SLR also emphasizes the transition from static, rule-based methods to adaptive and hardware-based solutions and identifies that the key trend is the shift toward adaptive cybersecurity mechanisms in heterogeneous IVN environments. Finally, the limitations in current solutions are also identified as research gaps, and the future research directions are outlined to strengthen the cybersecurity resilience of autonomous vehicles.
