Day 11 Dimensionality Reduction
Today I explored Dimensionality Reduction — a vital step to make sense of high-dimensional data 📉🔍
🔍 Analogy
Imagine packing for a trip. You have a massive wardrobe (your data), but your suitcase only fits 10 items. So, you pick versatile clothes — a few that cover most needs (formal, casual, warm, cold).
That’s Dimensionality Reduction: Compressing a huge dataset into its most meaningful parts — minimizing loss while maximizing utility.
🔹 Two Common Techniques
✅ PCA (Principal Component Analysis)
🧳 Like folding and layering smartly to save space, PCA finds directions (components) that capture the most variance (info).
Linear method
Great when data lies in neat, straight patterns
✅ t-SNE (t-distributed Stochastic Neighbor Embedding)
🧳 Like grouping clothes by outfits (shoes + formalwear), t-SNE clusters related data together.
Non-linear
Captures local relationships, distorts global structure
Ideal for visualizing complex, high-dimensional datasets
🚀 These techniques help models:
✅ Train faster
✅ Generalize better
✅ Reveal hidden patterns
🔐 Security Lens
⚠️ Information Loss & Blind Spots
🎒 You packed for summer, but forgot a raincoat. Rare threats (low variance) may get discarded — making your model blind to anomalies or attacks.
"Low variance" ≠ "low importance" — especially in security contexts.
⚠️ Feature Obfuscation by Attackers
🎒 Attackers can embed malicious patterns in dimensions likely to be discarded or compressed — bypassing detection pipelines.
⚠️ Inference Attacks on Embeddings
🎒 It's like sharing a blurred photo of your bag — someone could still guess your travel habits from the outline of items.
t-SNE visualizations can leak structural info — attackers might infer relationships between users, labels, or features. These compressed representations, if exposed, can be mined or reversed to extract sensitive patterns or identities.
📚 Key References
Jolliffe (2002) – Principal Component Analysis
Carlini et al. (2020) – Extracting Training Data from Embeddings
💬 Prompt
Have you visualized your model with t-SNE? What insights — or vulnerabilities — did you discover?
📅 Tomorrow
We explore KNN & Clustering — the simplest ML algorithms and how attackers exploit proximity logic 🔗👥
🔗 Missed Day 10?
Catch up here: https://lnkd.in/gMh3rr8b
#100DaysOfAISec - Day 11 Post #AISecurity #MLSecurity #MachineLearningSecurity #DimensionalityReduction #CyberSecurity #AIPrivacy #AdversarialML #LearningInPublic #100DaysChallenge #ArifLearnsAI #LinkedInTech
Last updated