UniShield: Forgery Detection & EMI Shielding

• UniShield is a dual-use technology featuring a multi-agent framework for unified forgery detection across diverse image manipulation domains, achieving state-of-the-art performance.
• The framework employs adaptive routing and specialized expert detectors to dynamically assess forgery types, ensuring high accuracy and cross-domain robustness.
• In materials, UniShield utilizes flexible polymer films embedded with quasi-one-dimensional van-der-Waals fillers for efficient, electrically-insulating EMI shielding.

UniShield denotes two technically distinct solutions unified by their focus on advanced protective mechanisms: (1) a multi-agent artificial intelligence framework for unified forgery image detection and localization across multiple image manipulation domains, and (2) a material technology based on flexible, electrically insulating polymer films with quasi-one-dimensional van-der-Waals conductor fillers for electromagnetic interference (EMI) shielding. Both usages exemplify domain-crossing unification: in the first, via adaptive model orchestration across disparate forgery types; in the second, by harnessing high-aspect-ratio van-der-Waals materials for scalable, robust electromagnetic shielding. Each implementation features precise engineering constraints, rigorous performance evaluation, and a pathway toward deployment in real-world security-relevant scenarios.

1. Forgery Detection and Localization: The UniShield Multi-Agent Framework

UniShield addresses the challenge of fully automated, cross-domain Forgery Image Detection and Localization (FIDL) by integrating specialist detection models under adaptive, agent-based orchestration. This is motivated by the increasing realism and societal risk of synthetic images, which amplify the need for scalable and practical FIDL solutions beyond domain-specific or monolithic approaches (Huang et al., 3 Oct 2025).

FIDL Problem Taxonomy

UniShield unifies four orthogonal forgery sub-domains:

• Image Manipulation Detection and Localization (IMDL): "Cheapfakes" such as splicing, copy–move, removal, and inpainting in natural images.
• Document Manipulation Detection and Localization (DMDL): Tampering in scanned documents (text, stamps, numbers).
• DeepFake Detection (DFD): Portrait forgeries (face-swap, reenactment).
• AI-Generated Content Detection (AIGCD): Fully synthetic images from GANs or diffusion models.

Existing detectors, while state-of-the-art within a single track, fail to generalize to others. Attempts to train a universal model—without agentic routing or adaptive mechanisms—produce “domain conflict” and overall degraded performance.

2. System Architecture and Workflow

UniShield comprises two principal collaborating agents and a summarizer module:

Component	Role	Notable Implementation
Perception Agent	Detects forgery domain and determines model flavor	Multimodal LLM (Qwen2.5-VL) with reinforcement learning
Detection Agent	Invokes the chosen expert detector, collects output	Eight state-of-the-art domain experts (LLM and non-LLM variants)
Summarizer	Structures interpretable output	GPT-4o, fixed prompt template for human-readable reports

The data flow is: Image → Perception Agent (Task Router → Tool Scheduler) → Detection Agent (Expert Detector) → Summarizer → Final Report.

Perception Agent

• Task Router: Assigns images to {IMDL, DMDL, DFD, AIGCD} using a fine-tuned multimodal LLM (Qwen2.5-VL), optimized via Guided Reinforcement with Prompt Optimization (GRPO). The objective incorporates both task reward (accuracy of domain assignment) and adherence to prompt formatting.
• Tool Scheduler: Decides (per track) between LLM-based (semantic inconsistency and high-level logic) versus non-LLM-based (low-level artifacts) expert invocation without further training, based on an auxiliary scan prompt.

3. Expert Detector Toolbox and Output

UniShield’s Detection Agent manages a toolbox with eight expert detectors, categorized by sub-domain and flavor:

Track	Non-LLM Expert	LLM-Based Expert
IMDL	IML-ViT (confidence + mask)	FakeShield (detection + mask + explanation)
DMDL	AscFormer (confidence + mask)	DMDL-R1 (detection + mask + explanation)
DFD	CLIP (confidence only)	DFD-R1 (detection + explanation)
AIGCD	AIDE (confidence only)	FakeVLM (detection + explanation)

Only one expert is deployed per image, based on the Perception Agent’s output, simplifying interpretability and avoiding conflicting cues. Each expert returns a scalar confidence score; IMDL/DMDL outputs include pixel-level masks; LLM-based experts add textual rationales for their decisions.

4. Training, Optimization, and Evaluation

• Task Router Training: GRPO RL is used to fine-tune Qwen2.5-VL for both domain routing and internal LLM-based expert design.
• Grounding Components: R1-V full supervision for DMDL-R1’s textual-to-mask translation (GLaMM).
• No End-to-End Backpropagation: Optimization is modular, with no joint pipeline-level gradient propagation.

UniShield’s evaluation spans multiple public datasets:

• IMDL: CASIA1+, IMD2020
• DMDL: RTM
• DFD: DF40 (FS and FR subsets)
• AIGCD: AIGCDetectionBenchmark

Key metrics: accuracy, F1 (image/pixel-level), IoU, AUC.

Empirical benchmarks demonstrate state-of-the-art performance across all sub-domains and substantial cross-domain robustness:

• DFD F1: 0.911 (vs. FakeShield’s 0.710)
• DMDL Img-F1: 0.736 (vs. ResNet’s 0.650)
• IMDL Img-F1: 0.96 (vs. FakeShield’s 0.95; IML-ViT’s 0.93)
• AIGCD mean acc.: 0.942 (AIDE: 0.928)

Ablation studies show dynamic, Perception Agent-driven expert scheduling outperforms fixed or ensemble voting strategies for both detection and localization.

5. Interpretability and End-User Reporting

Final results are packaged into a four-part report (via Summarizer/GPT-4o):

• Description: Scene summary
• Detection: Real vs. fake with confidence
• Localization: Tampering mask and free-form region description
• Judgment Basis: Bulletized cues (e.g., “Edge artifacts along spliced border,” “Inconsistent lighting”)

All intermediary agent outputs are formatted into GPT-4o prompts for human-readable but technically precise result interpretation.

6. Practical Implications, Adaptiveness, and Scalability

UniShield’s architecture supports:

• Adaptiveness: Automatic, perception-based routing and expert invocation obviate user-side specialization or prior knowledge.
• Scalability: New detectors or entire sub-domain tracks can be integrated without retraining the Perception Agent.
• Efficiency: Singular expert invocation reduces computational overhead; main bottleneck is MLLM inference, which is mitigatable through quantization or caching.

A plausible implication is broader suitability in forensic pipelines, journalistic content verification, and media rights enforcement, positioning UniShield as a unified FIDL guardrail for emerging misinformation risks (Huang et al., 3 Oct 2025).

7. UniShield as a Material Technology: Electrically-Insulating EMI-Shielding Films

The UniShield™ brand also encompasses a class of lightweight, flexible polymer films incorporating quasi-one-dimensional TaSe₃ van-der-Waals atomic thread bundles as EMI-screening fillers (Barani et al., 2021), offering:

• Material basis: High-aspect-ratio (AR up to $10^6$) TaSe₃ bundles (diameter 50–100 nm, length 200–500 μm) dispersed in UV-cured acrylate, epoxy, or sodium-alginate matrices.
• Growth and Exfoliation: TaSe₃ single crystals grown by chemical vapor transport at 750–650°C, exfoliated in acetone/DMF via sonication and centrifugation.
• Composite Structure: Uniform dispersion of atomic threads confirmed by SEM/TEM, phase/chemical stability by XRD/Raman/XPS.
• Shielding Performance:
  • X-band (8.2–12.4 GHz): SE$_\text{total}$ ≈ 10–20 dB for 1–3 vol% fillers, thickness 75–150 μm.
  • Sub-THz (220–320 GHz): SE$_\text{A}$ up to 76 dB for 1 mm thick, 1.3 vol% epoxy films.
  • DC conductivity: Remains below percolation threshold for $\varphi < 3$ vol%, ensuring electrical insulation.

Parameter	Flexible 5G/X-band	Module sub-THz/6G
Matrix	SA/UVP	Epoxy
TaSe₃ loading (vol%)	1.0–3.0	1.0–1.5
Thickness (μm/mm)	75–150 μm	0.5–1.5 mm
SE$_\text{total}$	10–20 dB @ 10 GHz	≥60 dB @ 300 GHz
Mass/Area	<1 g·cm$^{-2}$	<1 g·cm$^{-2}$

• Mechanical/Thermal Properties: Flexibility (>100 bending cycles, r ≥5 mm), tensile modulus 2–3 GPa, thermal stability >200°C, corrosion resistance (no oxidation below 300°C).
• Processing and Scale-Up: Compatible with high-shear mixing, roll-to-roll solvent coats, UV-lamp or mold-curing, allowing straightforward manufacturing scale-up.

The shielding mechanism is dominated by sub-threshold "antenna" action, where metallic thread bundles couple to incident E-fields and dissipate energy via reflection and absorption even without physical percolation, contrasting with conventional metal fill composites.

• "UniShield: An Adaptive Multi-Agent Framework for Unified Forgery Image Detection and Localization" (Huang et al., 3 Oct 2025)
• "Electrically-Insulating Flexible Films with Quasi-One-Dimensional van-der-Waals Fillers as Efficient Electromagnetic Shields" (Barani et al., 2021)