Day 7 Loss Functions
Loss Functions — The Model’s Moral Compass
Today I explored Loss Functions — the core ingredient that tells a machine learning model how “wrong” it is, and how to improve.
🔹 Quick Primer
A loss function measures the gap between predicted and true values
The goal of training is to minimize this loss
Models update their internal weights based on the loss to get closer to the truth
🧠 Think of a loss function as the moral compass of a model — but a poorly chosen or manipulated one can send it in the wrong direction.
🔐 Security Lens: How Loss Functions Can Be Attacked
⚠️ Loss Hijacking
Attackers subtly poison the training data so the loss function rewards the wrong behavior.
🦴 Analogy: You're training a dog to sit, giving treats. But someone gives treats when it jumps instead. Now the dog thinks jumping is “correct” — same with your model.
⚠️ Gradient Masking
Some defences modify loss functions to hide gradients — but attackers often find ways around them.
🧭 Analogy: You’re playing “hot and cold” to find treasure. Normally, “hot” means you’re close. But now someone fakes the sound. You keep hearing “hot” even when far — you’re misled into thinking you're doing well.
⚠️ Poor Loss Choices
Choosing the wrong loss function makes the model optimize the wrong thing.
🌶️ Analogy: You judge recipes only by how spicy they are, ignoring taste or balance. You’ll end up with a dish that's all spice, no flavor — a model good at the wrong goal.
📚 Key References
Goodfellow et al. (2014): Explaining and Harnessing Adversarial Examples
Carlini & Wagner (2017): Towards Evaluating the Robustness of Neural Networks
💬 Question for You
What’s the strangest model failure you’ve seen because of a poorly chosen loss function? Let’s discuss 👇
📅 Up Next: We dive into Gradient Descent — how models “walk downhill” to find better answers 🔽
🔗 Missed Day 6? Catch up here
#100DaysOfAISec – Day 7 Post #AISecurity #MLSecurity #MachineLearningSecurity #LossFunctions #CyberSecurity #AIPrivacy #AdversarialML #LearningInPublic #100DaysChallenge #ArifLearnsAI #LinkedInTech
Last updated