Unified Detection Paradigm

• Unified detection paradigms are end-to-end frameworks that integrate varied input types and modalities into a single model.
• They employ integrated representation learning, joint optimization, and statistical invariance to overcome the limits of siloed systems.
• These paradigms are applied in vision, radar, security, and anomaly detection to achieve state-of-the-art performance and robustness.

A unified detection paradigm refers to an architectural and algorithmic framework that enables detection systems to process, interpret, and make decisions about diverse input types, modalities, or adversarial conditions under a single, end-to-end model. These paradigms are specifically constructed to overcome the limitations of highly specialized or siloed systems—such as those that require separate models per data source, task, or attack—and instead centralize detection by leveraging cross-domain representations, joint architectural modules, and principled statistical or machine learning methodologies. Unified detection paradigms have seen influential developments in vision, radar, anomaly, and security domains.

1. Fundamental Principles of Unified Detection Paradigms

Core to the unified detection paradigm is the explicit design objective to create a single model or algorithmic pipeline capable of handling a spectrum of detection tasks or challenges, often in settings that previously required separate models or bespoke engineering. The main principles include:

• Heterogeneous Task/Modality Coverage: The system must process mixed types of data or tasks—e.g., physical and digital attacks in face spoofing (Fang et al., 31 Jan 2024), multiple sensor domains in 3D object detection (Zhang et al., 2023, Li et al., 28 Feb 2024), or open-vocabulary dense vision tasks (Tai et al., 10 Mar 2025).
• Integrated Representation Learning: Unified architectures employ representations (e.g., via joint token spaces, prompt-based modules, or multi-branch feature learning) that serve as a common base for varied downstream detection heads (Yang et al., 19 Nov 2025, Yang et al., 2022, Zhang et al., 2023).
• Joint Optimization and Loss Design: Training objectives enforce consistency, invariance, or cross-task performance, e.g., with cycle-consistency (Yang et al., 19 Nov 2025), cross-modal prompt fusion (Fang et al., 31 Jan 2024), or multi-task loss aggregation (Tai et al., 10 Mar 2025).
• Statistical or Invariance-Based Guarantees: In signal processing and radar, unification is often achieved at the algorithmic level using invariant statistics to attain constant false-alarm rate (CFAR) properties across a range of scenarios (Ciuonzo et al., 2015, Orlando et al., 2021, Addabbo et al., 2020, Zaimbashi et al., 26 Feb 2024).
• Extensible to Adversarial or Open-Set Scenarios: Advanced unified paradigms are robust to new, unseen attacks, dataset shifts, or adversarial patterns, leveraging, for example, prompt-based generalization or architecture-agnostic comparison metrics (Fang et al., 31 Jan 2024, Wang et al., 21 Mar 2025).

2. Key Architectures and Strategies

Unified detection systems span from data-level and representation-level integrations to modular architectures:

• Prompt-Based and Token-Level Unification: Vision-LLMs integrate both textual and visual context using teacher-student prompt modules, unified token spaces, or triplet-based decoding (Fang et al., 31 Jan 2024, Yang et al., 19 Nov 2025, Tai et al., 10 Mar 2025). For example, UniAttackDetection fuses CLIP-based teacher prompts (fixed templates for "real" vs "spoof") and student prompts (learnable vectors for real, physical, digital attacks) with a multi-branch feature mining module to handle all attack types under a strict two-class liveness regime (Fang et al., 31 Jan 2024).
• Query-Based or Embedding-Query Frameworks: In 3D point cloud understanding, the EQ-Paradigm decouples feature embedding (via any backbone) from query-by-location feature extraction, enabling flexible attachment to any detection head and arbitrary point-set inputs (Yang et al., 2022).
• Cross-Modality and Dataset Alignment Modules: Multi-dataset or multi-domain detectors deploy data-level corrections (e.g., range standardization, per-dataset normalization), semantic coupling-recoupling layers, and unified domain alignment to homogenize features across domains (Zhang et al., 2023, Li et al., 28 Feb 2024).
• Cycle-Consistent or Bidirectional Losses: Some frameworks employ bidirectional mappings and cycle consistency between input domains (e.g., images and HOI semantics (Yang et al., 19 Nov 2025)) or between detection and generation, improving generalization and semi-supervised learning.

3. Unified Paradigms in Security and Robustness

Modern unified detection also encompasses security-critical scenarios such as backdoor or adversarial trigger detection:

• Cross-Examination and Model Inconsistency: Lie Detector (Wang et al., 21 Mar 2025) pioneers a cross-examination approach: two independently trained models on (potentially tampered) outsourced data are compared via centered kernel alignment (CKA) of internal activations. A minimal input perturbation is optimized to create divergent behavior, exposing backdoors agnostic to model architecture or learning paradigm.
• Fine-Tuning Sensitivity: Unified detectors distinguish genuinely compromised models (backdoors) from merely adversarially perturbed or robust models by measuring the drop of attack success rate before and after fine-tuning on clean data.
• Limitations and Assumptions: Such security paradigms typically assume at least two uncoordinated service providers and some level of white- or grey-box access to model internals (Wang et al., 21 Mar 2025).

4. Statistical Unification in Detection Theory

Unified detection has deep roots in statistical signal processing, particularly for adaptive detection in radar and array processing:

• Maximal Invariant Statistics and CFAR: I-GMANOVA and related frameworks (Ciuonzo et al., 2015) construct detectors (GLRT, Rao, Wald, Gradient, etc.) as explicit functions of data-derived maximal invariants, guaranteeing CFAR across interference, clutter, and nuisance parameter regimes.
• Model Order Penalization: When considering multiple alternative hypotheses (e.g., multiple jammers, unknown target extent), the Kullback-Leibler Information Criterion (KLIC) paradigm unifies model selection and detection by penalizing log-GLRT scores by model dimensionality, preventing overfitting and ensuring adaptive CFAR (Addabbo et al., 2020).
• Unification over Domains and Signal Models: The unified theory of adaptive subspace detection (Orlando et al., 2021) prescribes different test statistics for first- or second-order subspace structures, each covering a broad spectrum of prior detectors as special cases; detector selection is thus unified by the nature of a priori information and environmental (homogeneity/heterogeneity) factors.

5. Empirical Impact and Performance Gains

Unified paradigms consistently surpass their siloed or per-task/model baselines:

• Face Anti-Spoofing: UniAttackDetection achieves state-of-the-art results under both physical (print/replay) and digital (deepfake, adversarial) attacks, with ACER as low as 0.52% and robust generalization to new, unseen attack types (Fang et al., 31 Jan 2024).
• Backdoor Detection: Lie Detector leads all prior baselines for detection success rate across supervised, self-supervised, and multimodal LLMs, while strictly controlling false positives via CKA-based cross-examination (Wang et al., 21 Mar 2025).
• 3D and Multi-Modality Detection: Uni3D and UniMODE report multi-point AP improvements compared to direct merging or dataset-specific detectors (Zhang et al., 2023, Li et al., 28 Feb 2024). Performance gains are attributed to normalization, semantic-level fusion, and domain alignment.
• HOI and Vision-Language: UniHOI outperforms prior approaches on long-tailed HOI detection (+4.9% rare mAP) and open-vocabulary generation (+42.0% interaction accuracy) (Yang et al., 19 Nov 2025). REF-VLM leverages triplet-based referring to match or exceed leading open-world detection methods (Tai et al., 10 Mar 2025).

6. Limitations and Open Problems

While unified detection paradigms deliver substantial flexibility and accuracy, current limitations are actively studied:

• Data and Label Coverage: Some paradigms (e.g., triplet-based referring in REF-VLM) structurally support new output formats (e.g., pose, normals) but are empirically limited by available task-annotated data.
• System Assumptions: Security paradigms may be compromised if independent providers collude or in settings where only black-box model access is possible (Wang et al., 21 Mar 2025).
• Scalability and Efficiency: The unification of extreme heterogeneous domains (e.g., remote sensing multi-sensor imagery to open-set anomaly in images/videos) remains a computational and data-annotation challenge.
• Generalization to Unseen Scenarios: While prompt-based and token-level unification improve open-set performance, extreme distributional shifts, rare classes, or subtle anomalies can still evade detection boundaries.

7. Representative Table: Key Modules and Domains in Unified Detection

Domain	Core Unification Mechanism	Representative Works
Vision	Prompt/Token, Cycle Consistency	(Fang et al., 31 Jan 2024, Yang et al., 19 Nov 2025, Tai et al., 10 Mar 2025)
3D/LiDAR	Query-Based Embedding, Normalization	(Yang et al., 2022, Zhang et al., 2023, Li et al., 28 Feb 2024)
Security	Model Cross-Examination, CKA	(Wang et al., 21 Mar 2025)
Statistical Radar	Maximal Invariant, CFAR	(Ciuonzo et al., 2015, Orlando et al., 2021, Addabbo et al., 2020)
Anomaly Detection	Layer-wise Queries, Masking	(You et al., 2022)

Unified detection paradigms thus provide a principled, extensible foundation for multi-task, cross-domain, and open-set detection, with architectures and loss functions expressly designed to guarantee statistical invariance, robust generalization, and high accuracy across a wide spectrum of practical drivers and adversarial factors.