Day 3 Regression vs Classification

Regression vs. Classification — And How Attackers Exploit Them

---

Day 03 Poster

Today I learned the core difference between the two foundational ML problems:

---

🔹 Regression

The model provides an output in numerical form. Example: You may want to predict how temperature changes as a plane gains altitude.

---

🔹 Classification

The model predicts which category the input belongs to (e.g., spam vs. not spam, fraud vs. legitimate). Example: The model predicts whether it's safe or unsafe to fly based on weather conditions.

These two problem types form the backbone of most real-world AI systems — and attackers know exactly how to exploit them.

---

🔐 Security Lens

⚠️ Classification Models

• Model Stealing Attacks Attackers send many inputs (emails, images, etc.) to a model's API and collect the output labels. They then train their own model using this data to replicate the original model’s behaviour — even without access to the original algorithm or training data.

• Adversarial Attacks Especially dangerous in image/text classification. Think: Stop signs tricking self-driving cars into interpreting them as speed limit signs.

---

⚠️ Regression Models

• Model Stealing Attacks These are often easier to reverse-engineer, because the outputs are detailed numerical values rather than simple categories. Example: If a model returns "$357,200" instead of just "expensive," an attacker can more accurately reconstruct the model’s decision boundaries.

• Inference Leaks An attacker inputs specific data points into the model and observes the confidence level of the output. If the model shows high confidence, it may suggest that the input was seen during training — a sign of membership inference leak.

💡 Example: A salary prediction model unintentionally exposing real people’s salaries. (This is especially relevant in deep learning tasks, where inputs are context- and attribute-rich.)

📝 Note: I refined the previous inference explanation to remove ambiguity.

---

📚 References

• (2014) Explaining and Harnessing Adversarial Examples

• (2016) Stealing Machine Learning Models via Prediction APIs

• PortSwigger Blog on ML Attacks

• Linear Regression blog by Aditya Bharathi

---

💬 Question

What are some misclassification risks you've seen in your industry (e.g., fraud, content filters, medical predictions)?

---

📅 Tomorrow: Exploring Overfitting & Underfitting — and how memorization leads to leakage.

📢 Follow along, share your favourite resources, or drop ideas below — I’m all ears!

🔗 Previous Post: Day 2

---

Day 3/100 — Beginning My #100DaysOfAISec Journey #AISecurity #MLSecurity #DeepLearning #CyberSecurity #100DaysChallenge #LearningInPublic #ArifLearnsAI