CyLens: Advanced Metasurface & CTI Systems

• CyLens is a family of technologies leveraging cylindrical symmetries to transform complex optical fields and orchestrate cyber threat intelligence.
• In metaphotonic applications, CyLens employs engineered metasurfaces with discretized phase profiles for diffraction-limited uniform focusing and high efficiency.
• In cyber defense, CyLens utilizes curriculum pre-trained LLMs in a cascaded, modular CTI pipeline to enhance threat attribution and remediation accuracy.

CyLens refers to a family of advanced systems and methodologies unified by their exploitation of cylindrical symmetries—either in optical wavefront manipulation via metasurfaces (metaphotonic CyLens) or in the agentic orchestration of cyber threat intelligence using LLMs (LLM-based CyLens). These distinct but homonymous technologies share a conceptual focus on transforming complex, high-dimensional data or fields into structured, high-utility outputs with precision and efficiency.

1. Principles and Definitions

In metaphotonics, CyLens denotes a cylindrical metasurface lens engineered to focus electromagnetic energy into a uniform line or circle by imposing a prescribed phase profile on the scattered or reflected wavefront. Architectures range from all-dielectric planar implementations for line focusing at optical frequencies (Ha et al., 2018), to cascaded, azimuthally-varying metasurfaces employing wave-matrix synthesis for arbitrary cylindrical field transformations (Lin et al., 2021).

In cyber defense, CyLens is a cyber threat intelligence (CTI) copilot leveraging curriculum-pre-trained, instruction-tuned LLMs integrated with modular natural language processing pipelines. The system orchestrates the entire CTI lifecycle—threat attribution, contextualization, detection, correlation, prioritization, and remediation—through a cascaded multi-agent and modular framework informed by a large, structured corpus of threat knowledge (Liu et al., 28 Feb 2025).

2. Design Methodologies and Theoretical Foundations

Metasurface CyLens

The metasurface CyLens achieves diffraction-limited, high-uniformity line focusing by encoding a 1D parabolic phase profile,

$\varphi(x) = k_0 (\sqrt{f^2 + x^2} - f),$

where $k_0 = 2\pi/\lambda$ is the free-space wavenumber, $f$ is the focal length, and $x$ parameterizes the aperture (Ha et al., 2018). The phase profile is discretized and mapped to local nanostructure geometry in the metasurface; for TiO$_2$ pillar arrays, the pillar radius is swept to tune the reflection phase continuously across $[0, 2\pi]$.

Cylindrical field synthesis via cascaded metasurfaces utilizes the wave-matrix formalism (Lin et al., 2021), where the total $z$-polarized electric field is expanded in cylindrical harmonics, and azimuthally-varying sheet admittances are synthesized via optimization of their Fourier components to enforce a desired modal output. Cascaded, multimode transmission matrices model complex inter-layer coupling and enable realization of arbitrary amplitude and phase profiles around a cylindrical contour.

LLM-based CyLens

The CyLens CTI copilot employs a three-phase pipeline:

1. Domain-specific pre-training on 271,570 CVE-centric threat reports—curated from MITRE-CVE, NVD, CWE, CAPEC, ATT&CK, and other incident repositories—using a staged pacing function to gradually expose model parameters to increasing complexity and diversity.
2. Instruction tuning on datasets for cascading, multi-stage CTI reasoning tasks.
3. Inference with modular NLP functions implementing topic modeling, NER, relation extraction, retrieval-augmented generation, reasoning, and summarization (Liu et al., 28 Feb 2025).

The architecture supports parallelizable module execution, context caching, and tool-style API wrappers, enabling autonomous, stateful orchestration of CTI workflows.

3. Performance, Evaluation, and Experimental Realization

Metasurface CyLens

Empirical characterization of the all-dielectric CyLens yields:

Parameter	Simulation	Experimental
Focal length $f$	400 μm	400 μm
Aperture $D$	200 μm	200 μm
Numerical aperture NA	0.247	0.247
Focus width $d$	1.6 μm	1.6 μm
Uniformity $k_0 = 2\pi/\lambda$0	0.99	0.92

The measured uniformity of 0.92 at 800 nm outperforms conventional glass cylindrical lenses (which typically achieve $k_0 = 2\pi/\lambda$1), as bulk optics are more affected by tolerance and alignment errors (Ha et al., 2018). The wave-matrix synthesized multilayer designs exhibit $k_0 = 2\pi/\lambda$2 focusing efficiency, FWHM $k_0 = 2\pi/\lambda$3 $k_0 = 2\pi/\lambda$4, and 3 dB bandwidth of $k_0 = 2\pi/\lambda$5–$k_0 = 2\pi/\lambda$6 (Lin et al., 2021).

LLM-based CyLens

Benchmarked against contemporary SOTA CTI agents and general-purpose LLMs, CyLens-8B and CyLens-70B models demonstrate:

• Threat actor identification accuracy: 87.6%/96.5% vs. 65% (GPT-4o).
• TTP F1: 92.2%/88.9% vs. 75–82%.
• CVSS Base accuracy: 96.9%/97.6% vs. 58–84% (Liu et al., 28 Feb 2025).

Zero-day generalization matches or exceeds historical performance in several descriptive and reasoning tasks. Ablation studies confirm the critical impact of curriculum pre-training and specialized NLP module integration, with removal of components (NER/REL/RAG) paring performance by 5–15%.

4. Implementation Details and Customization

Metasurface CyLens

Planar CyLens fabrication involves:

1. Substrate cleaning and multilayer stack deposition (Ge, Ag, SiO$k_0 = 2\pi/\lambda$7).
2. PMMA e-beam patterning and development.
3. Atomic layer deposition (ALD) of TiO$k_0 = 2\pi/\lambda$8 and RIE etch-back.
4. PMMA lift-off to finalize the metasurface.

Structural parameters (height, radius, periodicity) are selected to maximize phase coverage and to ensure unit-cell cross-sections align with the desired phase profile. Cascaded implementations require precise control over layer separation ($k_0 = 2\pi/\lambda$9) and unit-cell discretization to maintain modal fidelity (Lin et al., 2021).

LLM-based CyLens

CyLens is instantiated in three model scales (1B, 8B, 70B parameters), each deployable on common enterprise hardware (A10/A100 GPUs). Customization is facilitated by:

• Fine-tuning on organization- or domain-specific corpora.
• Prompt-engineering interfaces for document style and compliance requirements.
• Plugin APIs for integration with ticketing, SIEM, or vulnerability scanning systems.

Experiments demonstrate robust accuracy retention with as little as 50% target-specific data during adaptation phases (Liu et al., 28 Feb 2025).

5. Workflow, Applications, and Integration

Optical CyLens

Applications include:

• Hyperspectral Raman spectroscopy, leveraging the uniform, sub-micron line focus for enhanced volumetric signal.
• Planar line-focusing solar concentrators enabling wafer-scale PV modules.
• On-chip wave probes and biomedical scanners, exploiting the monolithic, CMOS-compatible form factor for integration (Ha et al., 2018).

LLM-based CyLens

Core CTI functionality encompasses the full threat management lifecycle:

1. Attribution (threat actor, TTPs, exploitation path).
2. Contextualization (affected assets, infrastructure, impact).
3. Detection (CVE/CWE identifiers).
4. Correlation (direct/intermediate CVE, CWE relationships).
5. Prioritization (CVSS, EPSS scoring/prediction).
6. Remediation (patch/tool recommendations, mitigation advisories) (Liu et al., 28 Feb 2025).

Modular pipeline execution facilitates parallel processing for latency reduction and reuse of intermediate context for multi-step queries.

6. Limitations and Considerations

Metasurface CyLens approaches are limited by achievable phase and amplitude modulation granularity, absorption losses (mitigated by low-loss dielectrics and wide metallic conductors), and phase-dispersion-induced bandwidth constraints (Ha et al., 2018, Lin et al., 2021). Layer separation and manufacturing precision are critical in cascaded implementations to avoid performance degradation.

In LLM-based CyLens, model predictions, particularly for zero-day EPSS trends, depend on real-time ingestion of new evidence, and performance may degrade in the absence of web access through the RAG module. Security and operational integrity are preserved via on-premise deployment, closed-network fine-tuning, and standard LLM safety mechanisms; however, reliance on web search (when enabled) introduces external dependencies (Liu et al., 28 Feb 2025).

7. Outlook and Significance

Metaphotonic CyLens architectures enable replacement of bulky, alignment-sensitive cylindrical optics with flat, integrable metasurfaces, directly coupling to modern optoelectronic and biosensing platforms. The wave-matrix synthesis formalism supports the design of arbitrary cylindrical scattering and focusing devices, advancing applications in imaging, energy concentration, and stealth technology (Ha et al., 2018, Lin et al., 2021).

The agentic, LLM-powered CyLens establishes a blueprint for autonomous, expert-level cyber threat analysis operating at web-data scale, supporting both commoditized and highly customized deployments. The integration of curriculum pre-training, cascading reasoning, and specialized NLP modules demonstrates the viability of LLMs as core components in the next generation of CTI workflows (Liu et al., 28 Feb 2025).