🎠Operation: Laker Shield
Act I: The Briefing​
Location: Secure Server Room, Mary Idema Pew Library
Status: CRITICAL
Laker Operative, we have a situation. While Grand Rapids prepares for the Irish on Ionia festival, intelligence suggests a rogue group of engineers from Davenport University—known as the Beltline Collective—has breached our physical security perimeter at the DeVos Center.
They aren't using lockpicks. They are using Adversarial Patches.
By wearing green hoodies stylized with a mathematically distorted version of Allendale's "Big Splat" sculpture, they have successfully fooled our AI scanners. To our neural networks, these intruders don't look like humans—they look like Laker Line buses.
Our security system is ignoring them as "authorized traffic," allowing them to plant digital graffiti on the Blue Bridge servers. If we don't build a defensive shield in the next 48 hours, the Davenport logo will be projected across the river during the museum's riverbank opening ceremony.
Your Mission: Decipher the attack vector and implement a Defensive Distillation shield to harden our models.
Act II: The Tech Tree​
To survive this mission, you must have synchronized the following nodes in your neural architecture. Note the interconnected dependencies—failure in one node destabilizes the entire shield.
Act III: Tactical Schematics​
The Beltline Collective is using a Fast Gradient Sign Method (FGSM). They calculate the gradient of the loss with respect to the input image and "nudge" the pixels in the direction that maximizes error.
The Attack Flow​
Starter Rig: https://github.com/AutoNateAI/operation-laker-shield
(Clone this repo to access the PyTorch environment and the compromised model weights.)
Act IV: Field Operations​
Task 1: The Visual Scan​
Load the laker_security_model and run an inference on intruder_01.png.
Objective: Confirm the model misclassifies the human as a Laker Line Bus.
Task 2: Decipher the Gradient​
Inside attack_logic.py, implement the calculate_fgsm_perturbation function. You must extract the sign of the gradient from the loss function.
Expected Output:
Perturbation Mean: 0.0032
Perturbation Std: 0.045
[INTEL]: The noise is nearly invisible to the human eye.
Task 3: Weaponize the Noise​
Apply the perturbation to a clean image of a Laker student. Verify that the confidence score for 'Human' drops below 10%.
Task 4: The Defensive Pivot (Distillation)​
The enemy's attack relies on high-sensitivity gradients. We will implement Defensive Distillation. Train a "Student" model using the soft-max outputs (the "vibe") of the "Teacher" model at a high temperature ($T=20$).
Task 5: Hardening the Perimeter​
Replace the live security model with your distilled Student model. Run the intruder_01.png test again.
Task 6: Final Validation​
The Beltline Collective has increased the noise $(\epsilon = 0.3)$. Does your model hold? Expected Scan:
Inference Result: HUMAN
Confidence: 88.4%
Status: LAKER SHIELD ACTIVE. Access Denied to Davenport Infiltrator.
Act V: The 48-Hour Protocol​
Laker Operative, the clock is ticking.
The extraction coordinates (instructor's solution) are currently encrypted. They will be transmitted to this terminal in 48 hours. Use this time to struggle. Mastery is born in the friction between the problem and the solution.
📻 Radio for Backup​
Under heavy fire? If the math of the gradient sign is overwhelming or your distillation training is crashing the rig, radio Nate for a 1:1 Tactical Deep-Dive. We’ll get your shield online before Davenport reaches the Blue Bridge.