BAAR Logo

BAAR4-guard Ultra (In-Line Multi-Era AI Threat Engine)

🌐 Official Website: https://baar.uk/model/
πŸ“¦ Model Hub: aixk/BAAR4-guard
πŸ“„ License: Non-Commercial Only (Contact for Commercial Pricing)


πŸ“Œ Model Overview (λͺ¨λΈ κ°œμš”)

BAAR4-guard Ultra is an ultra-fast, in-line edge AI threat detection and firewall engine trained across historical and modern traffic benchmarks spanning 2017 to 2026 (CIC-IDS2017, CICIoT2023, CICDDoS2019, and 2024~2026 Synthetic AI Vectors). Hardened with Multi-Stage PGD (Projected Gradient Descent) Adversarial Training, the model provides zero-trust packet/session classification with extreme resilience against adversarial payload perturbations and evasive botnets, delivering sub-millisecond inference on edge and WASM runtimes.

BAAR4-guard UltraλŠ” 2017~2026 μ „μ£ΌκΈ° μ‹€μ „ λ„€νŠΈμ›Œν¬ νŠΈλž˜ν”½(CIC-IDS2017, CICIoT2023, CICDDoS2019 및 2024~2026 μ΅œμ‹  AI 곡격 ν•©μ„± 벑터)을 톡합 ν•™μŠ΅ν•œ μ΄ˆκ²½λŸ‰Β·μ΄ˆμ €μ§€μ—° 인라인 인곡지λŠ₯ λ³΄μ•ˆ μ—”μ§„μž…λ‹ˆλ‹€. 특히 **닀단계 PGD μ λŒ€μ  ν›ˆλ ¨(Adversarial Training)**을 μ μš©ν•˜μ—¬ νŒ¨ν‚· λ³€μ‘° 및 AI 기반 우회 곡격(Adversarial Evasion)에 λŒ€ν•΄ 압도적인 λ°©μ–΄ 볡원λ ₯을 λ°œνœ˜ν•˜λ©°, μ˜€νƒ(FPR)을 μ΅œμ†Œν™”ν•˜λ©΄μ„œ 0.8ms μˆ˜μ€€μ˜ 인라인 검증 속도λ₯Ό 보μž₯ν•©λ‹ˆλ‹€.


πŸ›‘οΈ Defense & Performance Benchmarks (μ‹€μ „ λ°©μ–΄ μ§€ν‘œ)

Validation Setup: PGD Lv.4 Adversarial Perturbation Test | Hardware: PyTorch CUDA & Edge Runtime Export

Class / Category Classification Type Detection Rate (μ°¨λ‹¨μœ¨ / ν†΅κ³Όμœ¨) Error Metric (였λ₯˜μœ¨) Primary Threat Scenarios
Class 0: 정상 νŠΈλž˜ν”½ (Normal Human) Benign Flow 99.84% ν†΅κ³Όμœ¨ FPR: 0.16% 일반 업무 νŠΈλž˜ν”½, 정상 IoT 톡신, μΈκ°€λœ HTTP/S ν”Œλ‘œμš°
Class 1: AI μ •μ°° / 봇넷 (Recon/Botnet) Threat Flow 100.00% μ°¨λ‹¨μœ¨ FNR: 0.00% 자율 ν¬νŠΈμŠ€μΊ”, 취약점 μΈν…”λ¦¬μ „μŠ€ μŠ€μΊλ‹, C2 λΉ„μ½˜ 톡신
Class 2: μ›Ή 곡격 / DoS (Web Attack & DoS) Threat Flow 100.00% μ°¨λ‹¨μœ¨ FNR: 0.00% SYN Flood, ICMP Flood, DrDoS (UDP/NetBIOS/Portmap), Web Exploit
Class 3: ν¬λ¦¬λ΄μ…œ 곡격 (Credential Brute-force) Threat Flow 99.70% ~ 100.00% μ°¨λ‹¨μœ¨ FNR: <= 0.30% 계정 νƒˆμ·¨ μ‹œλ„, λΆ„μ‚° 브루트포슀, Credential Stuffing

⚑ μ’…ν•© λ°©μ–΄ 및 μ λŒ€μ  견고성 (Adversarial Robustness)

  • μ‹€μ „ μ’…ν•© 곡격 μ°¨λ‹¨μœ¨ (Attack Block Rate): 99.97% ~ 100.00% (미탐λ₯  FNR: 0.00% ~ 0.03%)
  • Lv.4 νŒ¨ν‚· λ³€μ‘° AI λ°©μ–΄λ ₯ (PGD Robustness): 99.11% ~ 99.58%
  • 정상 νŠΈλž˜ν”½ ν†΅κ³Όμœ¨ (Benign Pass Rate): 99.84% (μ˜€νƒλ₯  FPR: 0.16%)
  • μ΅œμ’… 검증 볡합 점수 (Overall Defense Score): 298.96 / 300.00

πŸ“Š Dataset & Threat Vector Distribution (ν•™μŠ΅ 데이터 ꡬ성)

전체 ν•™μŠ΅ 셋은 aixk/bg-1에 νŒ¨ν‚€μ§•λ˜μ–΄ λ™κΈ°ν™”λ˜μ—ˆμœΌλ©°, 총 31,333건의 μ—„μ„ λœ μ‹œκ³„μ—΄ μœˆλ„μš°/νŒ¨ν‚· ν”Όμ²˜λ‘œ κ΅¬μ„±λ˜μ—ˆμŠ΅λ‹ˆλ‹€:

🎯 [κ· ν˜• 빅데이터셋 μ™„μ„±] 총 31,333개 졜적 μ‹œκ³„μ—΄ μœˆλ„μš°
   β”œ 정상 ν”Œλ‘œμš° (Class 0) : 12,521건 (CIC-IDS2017 Tuesday + CICIoT2023 Benign + ν•©μ„± 데이터)
   β”œ AI 자율 μ •μ°° (Class 1) :  2,321건 (PortScan + 2024~2026 AI Recon Vectors)
   β”œ DoS / μ›Ή 곡격 (Class 2): 15,000건 (SYN/ICMP Flood, DrDoS UDP/Portmap/NetBIOS, Web Attacks)
   β”” ν¬λ¦¬λ΄μ…œ 곡격 (Class 3) :  1,491건 (Thursday Web Attack Brute-force + ν•©μ„± Stuffing)
  1. CIC-IDS2017 Benchmark:
    • Tuesday-WorkingHours (정상 5,312건)
    • Thursday-WorkingHours-Morning-WebAttacks (μ›Ή 침투/무차별 λŒ€μž… 1,806건)
    • Friday-WorkingHours-Afternoon-PortScan (λ„€νŠΈμ›Œν¬ μ •μ°° 4,213건)
  2. CICIoT2023 & CICDDoS2019 Large-Scale Vectors:
    • BenignTraffic, DDoS-ICMP_Flood, DDoS-SYN_Flood (CICIoT2023: 총 8,856건)
    • Syn, DrDoS_UDP, Portmap, DrDoS_NetBIOS (CICDDoS2019: 총 10,162건)
  3. 2024~2026 μ΅œμ‹  μœ„ν˜‘ μ£Όμž…κΈ° (Modern Injected Vectors):
    • 각 ν΄λž˜μŠ€λ³„ 4,000건씩 총 16,000건의 μ°¨μ„ΈλŒ€ λ‹€λ³€λŸ‰ λ³€μ‘° νŒ¨ν„΄ ν•©μ„± μ£Όμž…

⚑ Inference Speed & Runtime Benchmark (μΆ”λ‘  벀치마크)

2,000회 연속 단일 νŒ¨ν‚·/μœˆλ„μš° νŒλ³„ 벀치마크 κ²°κ³Ό:

Runtime Engine Single Inference Latency Model Binary Size Runtime Characteristics
ONNX Runtime (CPU/WASM) 0.882 ms 2.53 MiB 크둜슀 ν”Œλž«νΌ, μ—”λ“œν¬μΈνŠΈ Wasm ꡬ동 μ΅œμ ν™”
λΉ„μ••μΆ• .baar3 Direct Engine 1.788 ms 1.24 MiB Zero-Copy Direct-mmap λ°”μ΄λ„ˆλ¦¬, μ΅œμ†Œ λ©”λͺ¨λ¦¬ ν’‹ν”„λ¦°νŠΈ

μ΄ˆμ €μ§€μ—° 인라인 적용: 단일 μ„Έμ…˜ 뢄석 1ms 미만의 처리 μ„±λŠ₯을 μœ μ§€ν•˜μ—¬ λ„€νŠΈμ›Œν¬ 병λͺ© ν˜„μƒ μ—†λŠ” 투λͺ…ν•œ 인라인 L7/L4 λ°©ν™”λ²½ 톡합이 κ°€λŠ₯ν•©λ‹ˆλ‹€.


πŸ“¦ Export Formats & Deployment Artifacts (배포 파일 ꡬ성)

  • model.safetensors: 고속 λ‘œλ”© 및 λ³΄μ•ˆ 검증을 ν†΅κ³Όν•œ 순수 κ°€μ€‘μΉ˜ ν…μ„œ 파일
  • model.onnx: WebAssembly, C++, Go λ“± 에지 λŸ°νƒ€μž„ ν˜Έν™˜ 고속 ONNX κ·Έλž˜ν”„ (2.53 MiB)
  • model_guard_ultra_fp16.baar3: Zero-Copy Direct-mmap 기반 FP16 μ΄ˆκ²½λŸ‰ λ„€μ΄ν‹°λΈŒ 컀널 λ°”μ΄λ„ˆλ¦¬ (1.24 MiB)
  • config.json: 제둜 트러슀트 μž„κ³„μΉ˜ 및 μž…λ ₯ ν”Όμ²˜ Robust Scaler 메타데이터

πŸ“„ License & Commercial Terms (λΌμ΄μ„ μŠ€ 및 상업적 λ„μž… μ•ˆλ‚΄)

  • Non-Commercial Use Only (비상업적 μ—°κ΅¬Β·κ°œμΈ 이용 ν•œμ •):
    This model is strictly provided for non-commercial, research, and educational purposes only. Any commercial use without a separate commercial agreement is strictly prohibited.
    λ³Έ λͺ¨λΈμ€ 순수 연ꡬ, ꡐ윑 및 개인 λΉ„μ˜λ¦¬ λͺ©μ μ— ν•œν•΄μ„œλ§Œ 무료둜 μ œκ³΅λ©λ‹ˆλ‹€. 사전 ν—ˆκ°€ μ—†λŠ” 일체의 영리적 μ΄μš©μ€ μ—„κ²©νžˆ κΈˆμ§€λ©λ‹ˆλ‹€.

  • Commercial Licensing & Inquiries for Revenue-Generating Entities (맀좜 λ°œμƒ κΈ°μ—… 및 μƒμš©ν™” λ„μž… 문의):
    If your organization generates revenue, or if you plan to incorporate this model into commercial products, services, or paid APIs, you are required to purchase a commercial license. Please contact us directly via email for enterprise licensing and integration support.
    κΈ°μ—…/사업체에 맀좜이 λ°œμƒν•˜κ³  μžˆκ±°λ‚˜, λ³Έ λͺ¨λΈμ„ μƒμš© μ„œλΉ„μŠ€Β·μœ λ£Œ μ œν’ˆΒ·μˆ˜μ΅ 창좜 λͺ©μ μ˜ μ‹œμŠ€ν…œμ— λ„μž…ν•˜κ³ μž ν•˜μ‹œλŠ” 경우 λ°˜λ“œμ‹œ λ³„λ„μ˜ μƒμš© λΌμ΄μ„ μŠ€λ₯Ό μ·¨λ“ν•˜μ…”μ•Ό ν•©λ‹ˆλ‹€. λ„μž… 쑰건 및 κΈ°μ—…μš© λΌμ΄μ„ μŠ€ λ°œκΈ‰μ€ μ•„λž˜ μ΄λ©”μΌλ‘œ λ¬Έμ˜ν•΄ μ£Όμ‹œκΈ° λ°”λžλ‹ˆλ‹€.

    • λ„μž… 및 μƒμš© λΌμ΄μ„ μŠ€ 문의 (Commercial Inquiries): admin@099.kr
    • 곡식 μ›Ήμ‚¬μ΄νŠΈ (Official Website): https://baar.uk/model/

⚠️ Notes & Disclaimer (기술 κ³ μ§€ 및 μ•ˆλ‚΄)

  • Defensive Engineering & Evaluation: λ³Έ λͺ¨λΈμ€ 인라인 μΉ¨μž… λ°©μ§€ μ‹œμŠ€ν…œ(IPS) 및 제둜 트러슀트 μ•„ν‚€ν…μ²˜ 연ꡬλ₯Ό μœ„ν•΄ μ œμž‘λ˜μ—ˆμŠ΅λ‹ˆλ‹€. μ‹€μ œ 운영 λ„€νŠΈμ›Œν¬μ— 배치 μ‹œ, 쑰직의 λ³΄μ•ˆ 정책에 맞좰 μ˜€νƒ ν—ˆμš©μΉ˜μ™€ 차단 μž„κ³„κ°’(config.json)을 사전에 μ‘°μ •ν•˜μ—¬ μ μš©ν•˜μ‹­μ‹œμ˜€.

πŸ’Ό Opportunities & Contact (μ±„μš© μ œμ•ˆ 및 νŒŒνŠΈλ„ˆμ‹­)

This project demonstrates end-to-end competency in adversarial network traffic defense, multi-source dataset engineering, PGD adversarial training, and ultra-low latency inference packaging (.baar3, ONNX).

I am actively seeking AI Engineering / Security AI Research opportunities, recruitment offers, and technical partnerships.

  • 곡식 μ›Ήμ‚¬μ΄νŠΈ (Official Website): https://baar.uk/model/
  • κΈ°μ—… λ„μž… 및 λΌμ΄μ„ μŠ€ 문의 (Commercial Adoption): admin@099.kr
  • μ±„μš© 및 ν¬μ§€μ…˜ μ œμ•ˆ (Recruitment & Hiring): admin@099.kr
  • 투자 및 기술 ν˜‘λ ₯ 문의 (Investment & Partnerships): admin@099.kr

극단적인 λ¦¬μ†ŒμŠ€ ν™˜κ²½μ—μ„œλ„ 고속 좔둠이 κ°€λŠ₯ν•œ λ”₯λŸ¬λ‹ μ•„ν‚€ν…μ²˜λ₯Ό μ„€κ³„ν•˜κ³ , μ‹€μ „ 사이버 μœ„ν˜‘ 데이터 μœ΅ν•© νŒŒμ΄ν”„λΌμΈλΆ€ν„° μ λŒ€μ  ν›ˆλ ¨ 및 μ–‘μžν™” λ°”μ΄λ„ˆλ¦¬ λ°°ν¬κΉŒμ§€ μ „ 과정을 직접 κ΅¬ν˜„ν•˜λŠ” AI μ—”μ§€λ‹ˆμ–΄μž…λ‹ˆλ‹€. 기술 λ„μž…, μ±„μš© μ œμ•ˆ 및 ν˜‘μ—…μ— 관심 μžˆλŠ” κΈ°μ—…κ³Ό νŒ€μ˜ 문의λ₯Ό ν™˜μ˜ν•©λ‹ˆλ‹€.

Downloads last month
-
Safetensors
Model size
644k params
Tensor type
F32
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Collection including aixk/BAAR4-guard