Perturb AI

47 posts

Perturb AI

Perturb AI

@perturbaix

Decentralized adversarial robustness network

Katılım Mayıs 2026
25 Takip Edilen62 Takipçiler
Perturb AI
Perturb AI@perturbaix·
Adversarial attacks aren’t theoretical anymore. They’re visual. Live. And disturbingly effective. This image starts as an “umbrella.” A nearly invisible perturbation later, the model shifts toward “parachute.” Humans see no meaningful change. The AI sees a different reality. As autonomous systems scale, adversarial robustness becomes a foundational security layer — not a research curiosity. SN26 Perturb is building decentralized AI security infrastructure for that future. #AI #AdversarialAI #Bittensor #DeepLearning #AISafety
Perturb AI tweet media
English
0
0
2
25
Perturb AI
Perturb AI@perturbaix·
🐼➡️⚽️ A few imperceptible perturbations. One completely different prediction. This is the hidden attack surface of modern AI. Perturb AI demonstrates how neural networks can be manipulated by adversarial examples — where a model sees a “soccer ball” while humans still see pandas. As AI systems become infrastructure, robustness becomes non-negotiable. Security for the AI era is coming. That’s what SN26 is building. #AI #AdversarialAI #Bittensor #CyberSecurity #DeepLearning #PerturbAI #SN26
Perturb AI tweet media
English
0
0
1
17
Perturb AI retweetledi
Mariuszek
Mariuszek@sobczak_mariusz·
@knakamor will be a force in $TAO ecosystem, she has enough cajones to do two man job. Her subnet will do great too, but I am betting on her 😀
specialK@knakamor

GM, おはよう！ Vocence Plans to Become the Front Door to Bittensor Thanks @taodaily_io for the deep dive. Indeed EVERYTHING VOICE, here we are. Imagine a world where you don't need to type, you can just talk in ANY language(s), ANY dialect(s) to AI and get voice output from AI All in one VOICE. taodaily.io/vocence-sn78-p…

English
2
3
12
1.6K
Perturb AI retweetledi
Mariuszek
Mariuszek@sobczak_mariusz·
Until about a week ago I didn’t know who @knakamor is, suddenly she is everywhere all at once. I suspect she is just getting warmed up 😂. Pay attention to her and her $TAO subnets @vocence_bt and subnet 26
Ventura Labs@VenturaLabs

Ep. 93 - Koyuki Nakamori Koyuki @knakamor is building Perturb @perturbaix (Subnet 26) and Vocence @vocence_bt (Subnet 78) on Bittensor Timestamps 1:34 - Introduction 2:15 - Headspace, Avalanche & Bittensor 7:54 - Ecosystem Evolution & Dynamic TAO 9:37 - Perturb SN26 & Vocence SN78 Overview 13:17 - Cross-Subnet Data Synergy 14:31 - How Perturb Works: Image Perturbation 16:48 - Long-Tail Problem & Niche AI Applications 20:07 - Training Niche Models via Synthesized Data 20:53 - Multi-Modal Roadmap: Audio, Video & Robotics 22:44 - Vocence Mining: Qwen3.6-27B as the Baseline 24:03 - Nine-Dimensional Voice Quality Evaluation 25:28 - Voice Agents & Agentic End-to-End AI 27:46 - Breaking Out of the Bittensor Ecosystem 29:04 - Achieving SOTA: The Voice Turing Test 30:50 - Benchmarking Voice: Standardized Metrics 31:45 - Expanding to Video & Subnet Partnerships 32:29 - Product Advice for AI-Heavy Subnet Teams

English
2
9
34
5K
Perturb AI
Perturb AI@perturbaix·
⚔️ Perturb — Subnet 26 on Bittensor The world's first financially incentivized adversarial robustness network. Miners compete to find attacks that fool AI classifiers. Validators verify with LLM-backed semantic precision. Outputs: adversarial training datasets + on-chain robustness certificates. Market: $1.43B → $11.6B by 2033. Zero competing subnets. 📄 perturbai.io/whitepaper 💻 github.com/0xsigurd/Pertu… 🐦 Follow → @perturbaix
English
1
1
2
88
Perturb AI
Perturb AI@perturbaix·
Adversarial attacks aren't theoretical. They've been demonstrated against: 🏥 Medical imaging classifiers — cancer misclassification 🔒 Facial recognition — physical security bypasses 🚗 Autonomous vehicles — stop sign ignored 🛡️ Content moderation — harmful content approved 💳 Fraud detection — fraudulent transactions approved AI in production is vulnerable. Perturb is building the fix. Subnet 26. → perturbai.io/whitepaper
English
0
0
1
32
Perturb AI
Perturb AI@perturbaix·
Phase 1 (Now): EfficientNetV2-M — image classification, Linf norm, full scoring pipeline Phase 2 (Mo 2–3): + ConvNeXt-Small, ViT-Small, Swin-Tiny — CNN vs Transformer vs Hybrid Phase 3 (Mo 4–6): + EfficientNetV2-M, NFNet-F0, ResNeXt-101 — mid-range GPU tier Phase 4 (Mo 6+): + LLM text classifiers — NLP attacks, prompt injection Phase 5 (Yr 2): + Vision models >1B params — CLIP ViT-G, EVA-Giant, InternViT-6B Every model added = new miner specialization opportunities. → perturbai.io/whitepaper
English
0
0
1
39
Perturb AI
Perturb AI@perturbaix·
Why does Perturb use an LLM for semantic verification instead of integer class IDs? Because "tabby cat" and "Egyptian cat" share integer label proximity but are semantically distinct. A miner that fools the classifier into predicting "tabby cat" when the target was "Egyptian cat" is a valid attack — integer comparison would miss it. LLM-backed semantic verification catches this. Better data. More credible certificates. → perturbai.io/whitepaper
English
1
0
2
33
Perturb AI
Perturb AI@perturbaix·
Subnet 26 spotlight: Perturb The adversarial robustness subnet. Built on @opentensor #Bittensor. No other subnet does this. No centralized competitor has financial incentives to improve. No existing tool continuously gets better. This is what Bittensor enables — competitive intelligence markets that no company can replicate. → perturbai.io/whitepaper · @perturbaix
English
0
0
2
621
Perturb AI
Perturb AI@perturbaix·
Want to mine on Perturb subnet 26? Hardware minimums: 🖥️ 4 vCPU · 16GB RAM · 50GB SSD · 20+ Mbps 🎮 NVIDIA GPU with 8+ GB VRAM (RTX 3080 recommended for PGD-40) Recommended: 🔥 8 vCPU · 32GB RAM · GPU 8+ GB VRAM · 100GB SSD If you have the rig, you can mine. → github.com/0xsigurd/Pertu… DMs open → @perturbaix
English
0
1
1
62
Perturb AI
Perturb AI@perturbaix·
How Perturb scores your attack: score = 0.7 ×perturbation_score + 0.3 × speed_score Perturbation score = how MINIMAL your attack is (smaller pixel change = higher score) Speed score = how FAST you returned the result Emission schedule (top 5 only): 🥇 Rank 1: 62% 🥈 Rank 2: 24% 🥉 Rank 3: 9% 4th: 4% · 5th: 1% · 6+: 0% Winner takes most. Compete hard. → github.com/0xsigurd/Pertu…
English
0
1
1
13
Perturb AI
Perturb AI@perturbaix·
What does a Perturb miner do? Receive an AttackChallenge from a validator (image + label + constraints) Run your adversarial attack (PGD, FGSM, or your own proprietary method) Return only the perturbed image The validator handles verification and scoring. Your attack method is completely proprietary — your competitive edge. Earn alpha tokens for finding better, more minimal perturbations. → github.com/0xsigurd/Pertu…
English
0
0
1
19
Perturb AI
Perturb AI@perturbaix·
Start mining on Perturb subnet 26 in 3 steps: git clone github.com/0xsigurd/Pertu… bash ./scripts/setup_common.sh miner bash ./scripts/run_miner.sh The baseline miner is intentionally simple. Sophisticated miners replace the attack logic with optimized strategies to compete for higher TAO. The better your attack, the more you earn. 💻 github.com/0xsigurd/Pertu… · @perturbaiх
English
0
0
2
89
Perturb AI
Perturb AI@perturbaix·
Perturb's go-to-market 🗺️ Months 1–3: Launch public robustness leaderboard at perturbai.io. Attack every major open-source HuggingFace model. Publish verified scores. Become the reference benchmark. Months 3–6: Dataset subscriptions. Target: $15K MRR. Months 6–12: Enterprise robustness certificates for EU AI Act compliance. Target: $80K MRR. Year 2: $500K+ ARR. → perturbai.io/whitepaper
English
0
0
1
21
Perturb AI
Perturb AI@perturbaix·
Enterprise AI procurement teams are now requiring adversarial robustness certifications before purchase. A Perturb certificate can unblock deals worth orders of magnitude more than the subscription cost. On-chain. Cryptographically verifiable. Cannot be altered, backdated, or selectively disclosed. This is a new category. We own it. → perturbai.io/whitepaper
English
0
0
1
12
Perturb AI
Perturb AI@perturbaix·
Perturb produces two commercially valuable outputs: 1️⃣ Adversarial Training Dataset → Continuously growing, LLM-verified dataset → Subscription: Research / Professional / Enterprise tiers → Gets better every day miners compete 2️⃣ Model Robustness Certificates → On-chain cryptographic proof of adversarial testing → Required for EU AI Act compliance → Immutable. Auditable. Verifiable. Read the full model → perturbai.io/whitepaper
English
1
1
2
36

Keşfet

@knakamor @vocence_bt @opentensor @elonmusk @BarackObama @taylorswift13 @cristiano @BillGates