DEV Community: nasihuyxbigmo

Beyond the Black Box: Securing Trust in Medical AI with Google Gemini

nasihuyxbigmo — Tue, 03 Mar 2026 22:39:42 +0000

This is a submission for the Built with Google Gemini: Writing Challenge

Beyond the Black Box: Securing Trust in Medical AI with Google Gemini

What I Built with Google Gemini

In a recent independent research sprint, I engineered an explainable multi-class lung cancer classification system using transfer learning and adversarial robustness analysis.

The system classifies histopathology images into:

Adenocarcinoma
Squamous Cell Carcinoma
Normal tissue

The backbone architecture is EfficientNetB0 (ImageNet-pretrained), fine-tuned with:

Selective unfreezing of upper convolutional layers
Cosine Decay learning rate scheduling
Stratified K-Fold Cross Validation
Macro F1-Score, Precision-Recall Curve, and ROC-AUC evaluation
Grad-CAM for visual interpretability
Preliminary adversarial perturbation testing

This was not an accuracy-focused experiment. It was an attempt to answer a more difficult question:

How can we make medical AI not only accurate — but explainable, robust, and secure?

In clinical environments, black-box predictions introduce cognitive and operational risk. A system that predicts malignancy without providing interpretability cannot support accountable decision-making. In medical diagnostics, opacity is not merely a technical limitation — it is a liability. Therefore, this project was designed from the beginning to integrate explainability and security considerations into the modeling pipeline.

Beyond raw classification performance, I focused on reducing false negatives, since missing a malignant case is significantly more critical than misclassifying a benign one. This required careful metric selection and threshold analysis rather than relying on overall accuracy.

The Role of Google Gemini

Google Gemini functioned as a structured reasoning partner rather than a code generator.

It helped me:

Compare architectural trade-offs between ResNet-50, MobileNetV2, and EfficientNetB0 (FLOPs, parameter efficiency, generalization behavior in small medical datasets)
Analyze gradient flow during Grad-CAM implementation when tensor shape mismatches occurred
Refine evaluation strategy beyond accuracy, emphasizing Macro F1-Score and false negative sensitivity
Structure adversarial robustness testing hypotheses

Instead of replacing engineering judgment, Gemini accelerated hypothesis-driven experimentation.

For example, when validation loss oscillated while training loss decreased, I did not ask a generic debugging question. I structured the prompt with architectural configuration, optimizer choice, and learning rate schedule. Gemini responded with possible explanations related to overfitting, batch normalization instability, and data variance — enabling faster experimental iteration.

In this way, Gemini served as a cognitive multiplier rather than an automation shortcut.

Demo

This project currently exists as a research-grade prototype pipeline with the following stages:

Image normalization and augmentation
EfficientNetB0 fine-tuning
Grad-CAM heatmap generation
Adversarial noise sensitivity testing (FGSM-style perturbation)
Confusion matrix and PR-curve evaluation

The pipeline begins with stain-normalized preprocessing to reduce domain shift. After feature extraction and fine-tuning, Grad-CAM overlays are generated to visualize discriminative regions influencing model decisions. To test robustness, small adversarial perturbations are injected to evaluate prediction stability under noise conditions.

The next milestone includes benchmarking against Vision Transformers (ViT) and preparing a journal-ready manuscript aligned with SINTA 2/3 publication standards. The long-term objective is not only to improve performance but to quantify explainability effectiveness and robustness under adversarial scenarios.

What I Learned

1. Transfer Learning Is Not a Silver Bullet

Pretrained models accelerate convergence but do not guarantee generalization.

Without:

Careful layer unfreezing
Learning rate scheduling
Regularization control

EfficientNet overfit aggressively on histopathology data.

Pretrained weights are initialization — not validation.

I learned that blindly trusting pretrained architectures leads to fragile systems. Fine-tuning required experimentation with freezing depth, adjusting weight decay, and carefully monitoring validation behavior. Optimization discipline mattered more than model complexity.

2. Explainability Is a Defensive Mechanism

Grad-CAM exposed instances of shortcut learning, where the model focused on staining artifacts rather than cellular morphology.

This transformed explainability from a visualization feature into an auditing layer.

If heatmaps highlight irrelevant structures, the model is not clinically trustworthy.

Explainability revealed when the network relied on spurious correlations instead of pathological features. In this sense, Grad-CAM acted as a diagnostic tool for the model itself. Interpretability became a safeguard against hidden bias and unintended feature exploitation.

3. Robustness Is as Important as Accuracy

With a cybersecurity background and prior research in DNA-based image cryptography, I evaluated the model through a threat lens.

If small perturbations can flip a cancer diagnosis, the system is unsafe.

This led to exploration of:

Adversarial perturbation sensitivity
Conceptual cryptographic integrity validation before inference
Secure deployment strategies

Medical AI must integrate:

Statistical learning
Interpretability
Cryptographic integrity
Adversarial robustness

It is not purely a machine learning challenge. It is a security engineering problem.

In particular, my prior research in image cryptography shaped my understanding of data integrity risks. If a histopathology image is tampered with at the pixel or bitstream level, both prediction and explanation can be corrupted. Therefore, model robustness must be complemented by data authenticity verification.

4. Prompt Precision Determines AI Quality

When prompts were vague, responses were generic.

When prompts included:

Architecture configuration
Observed loss behavior
Hypothesis framing

Gemini delivered structured, analytical reasoning.

AI amplifies thinking depth. It does not replace it.

This experience reinforced a fundamental insight: the quality of AI output directly reflects the clarity of the problem definition. Structured prompts transformed Gemini from a conversational assistant into a research collaborator capable of meaningful technical dialogue.

Google Gemini Feedback

What Worked Well

Strong contextual reasoning in architecture comparison
Clear explanation of backpropagation mechanics
Effective support in structuring evaluation methodology
Fast iteration during debugging cycles

Gemini was most powerful when treated as a research collaborator rather than an automated code tool. Its strength lies in accelerating structured reasoning and enabling rapid exploration of design alternatives.

Where Friction Occurred

Occasional hallucination regarding specific library versions
Overconfident answers without explicit uncertainty
Generic first responses when prompts lacked technical specificity

These limitations reinforced the necessity of independent verification. AI can suggest directions, but correctness must always be validated experimentally and technically.

This reinforced a critical lesson:

AI accelerates development, but verification and accountability remain human responsibilities.

Closing Reflection

This project reshaped how I view AI engineering.

The future is not AI replacing engineers.

The future is engineers who understand:

Statistical optimization
Model interpretability
Adversarial risk
Cryptographic data integrity

Google Gemini sharpened my reasoning and accelerated iteration cycles.

But methodological rigor, validation discipline, and security awareness remained mine.

And that is precisely how AI should be used.

Beyond the Black Box: Securing Trust in Medical AI with Google Gemini

nasihuyxbigmo — Tue, 03 Mar 2026 22:30:15 +0000

1. The Build: Designing a High-Stakes Histopathology Intelligence System

During a recent independent research sprint, I engineered an explainable multi-class lung cancer classification system (Adenocarcinoma, Squamous Cell Carcinoma, and Normal tissue) using transfer learning based on EfficientNetB0.

This was not a Kaggle-style accuracy chase.

It was an attempt to answer a harder question:

How do we make medical AI not only accurate — but trustworthy, auditable, and secure?

In clinical environments, a black-box model is a liability. A prediction without interpretability introduces cognitive risk. If a system flags a biopsy as malignant but cannot justify its reasoning, the physician inherits epistemic uncertainty without support.

Therefore, the system I built combined:

EfficientNetB0 backbone (ImageNet-pretrained)
Fine-tuned upper convolutional blocks
Stratified K-Fold Cross Validation
Macro F1-Score, Precision-Recall Curve, and ROC-AUC evaluation
Grad-CAM for visual interpretability
Preliminary adversarial robustness testing

The objective was not merely performance — it was clinical-grade accountability.

2. The Role of Google Gemini: From Assistant to Research Multiplier

Google Gemini was not used as a code generator.

It was used as a structured reasoning partner.

Architectural Trade-Off Analysis

One of the first design decisions involved selecting the backbone architecture. I explored:

ResNet-50 (deeper residual learning)
MobileNetV2 (lightweight edge-oriented model)
EfficientNetB0 (compound scaling optimization)
Rather than relying on intuition, I used Gemini to dissect trade-offs:
FLOPs vs parameter efficiency
Overfitting tendencies in small medical datasets
Feature resolution preservation in histopathology

Histopathological images contain high-frequency cellular textures. Overly aggressive downsampling can erase diagnostically relevant microstructures. Gemini helped structure that reasoning before I even ran experiments.

This shifted the workflow from trial-and-error to hypothesis-driven experimentation.

Deep Debugging: Grad-CAM Tensor Failure
During Grad-CAM implementation, I encountered tensor shape mismatch when extracting gradients from the final convolutional block.

Instead of asking “Why is this broken?”, I structured the question:

“Given EfficientNetB0 with top layers unfrozen, and Grad-CAM computed on the last convolutional block, why would gradient dimensions mismatch during backpropagation if the model includes GlobalAveragePooling2D before dense classification?”

That specificity changed everything.

Gemini responded by explaining:

How gradients flow backward from softmax output
The necessity of intercepting activations before pooling layers
Why spatial feature maps must remain intact for proper localization

This was not patch-level debugging.
It was conceptual reinforcement of how CNN interpretability actually works.

Research-Driven Evaluation Strategy

Gemini also challenged my initial evaluation design.

Accuracy is misleading in medical AI.

With class imbalance and high cost of false negatives, Macro F1-score and Precision-Recall analysis become more meaningful. Gemini helped me refine:

Stratified K-Fold Cross Validation to mitigate leakage
Confusion matrix analysis targeting false negatives
Threshold calibration strategies

In this context, AI did not replace methodological rigor — it accelerated it.

3. The Cybersecurity Perspective: Where ML Meets Adversarial Risk

My background extends beyond machine learning into cybersecurity research, including involvement in university-level cyber defense initiatives and research in DNA-based image cryptography.

That lens fundamentally altered how I evaluated this system.

Model Robustness Against Adversarial Attacks

A cancer classifier vulnerable to adversarial perturbation is dangerous.

If imperceptible pixel-level noise can flip a diagnosis, then the system cannot be trusted in real clinical pipelines.

Using Gemini as a conceptual guide, I explored:

FGSM-style perturbations
Sensitivity of EfficientNet feature maps to structured noise
The relationship between adversarial vulnerability and shortcut learning

What emerged was clear:

High accuracy ≠ high robustness.

Explainability revealed when the model relied on staining artifacts or slide edges — classic shortcut learning behavior.

Grad-CAM became more than visualization.
It became a forensic auditing layer.

Data Integrity: Cryptography Meets Medical AI

My prior research in DNA-based image cryptography directly influenced this project’s threat model.

Medical images must preserve:

Integrity
Authenticity
Confidentiality

If histopathology images are tampered at the bitstream level, Grad-CAM visualizations can become misleading. An attacker could alter texture patterns subtly, influencing both prediction and explanation.

This creates a compounded risk:

Manipulated input → Corrupted inference → False explanation → Clinical misjudgment.

To mitigate this conceptual vulnerability, I investigated:

Hash-based integrity verification pipelines
Pre-classification cryptographic validation
Secure image storage models

This intersection — between cryptography and explainable AI — is rarely discussed.

But it is necessary.

Medical AI is not just a machine learning problem.
It is a security engineering problem.

4. Lessons: Engineering Discipline Over AI Hype
Transfer Learning Is Not a Silver Bullet

Without careful unfreezing strategy and cosine decay learning rate scheduling, EfficientNet overfit aggressively.

Pretrained weights provide initialization — not guaranteed generalization.

Fine-tuning required:
Layer-wise learning rate control
Regularization

Careful monitoring of validation loss oscillation

Gemini accelerated iteration, but the responsibility of experimental discipline remained mine.
**
Explainability as a Defensive Mechanism**

Grad-CAM exposed weaknesses in model reasoning.

When heatmaps highlighted irrelevant slide regions, it revealed shortcut learning.

That insight was not cosmetic.

It was a security signal.

Explainability is not decoration — it is audit infrastructure.

Prompt Engineering Mirrors Cognitive Depth

When I asked vague questions, I received shallow answers.

When I framed hypotheses, constraints, and architecture details precisely, Gemini responded with structured reasoning.

This was a critical realization:

AI amplifies your thinking level.
It does not compensate for its absence.

5. Honest Feedback on Google Gemini
Strengths

Strong contextual reasoning in ML architecture comparison
Clear explanation of optimization strategies
Efficient assistance in modularizing preprocessing pipelines

Helpful in drafting structured research arguments

Gemini excelled when treated as a collaborator in reasoning — not as a code vending machine.

Limitations

Occasional hallucination in specific library version references
Overconfident suggestions without uncertainty disclaimers
Generic first responses if prompts lacked technical structure

The solution was not distrust.

The solution was supervision.

AI does not eliminate the need for engineering competence.
It increases the cost of intellectual laziness.
**

Roadmap: Toward Publication and Secure Deployment**

This project is not an endpoint.

It is evolving toward:

Publication (SINTA 2/3 Target And Scopus Target)

A formal research manuscript focusing on:

Grad-CAM effectiveness in reducing false negatives
Comparative benchmarking: CNN vs Vision Transformers
Robustness evaluation under adversarial perturbations Advanced Model Benchmarking

Testing Vision Transformers (ViT) to evaluate whether self-attention mechanisms capture global tissue patterns more effectively than convolutional local features.

Secure Deployment

Conversion to TensorFlow Lite / ONNX
Edge deployment under encrypted storage
Cryptographic verification before inference
Model integrity monitoring

This bridges:

Machine Learning
Explainability
Cryptography
Cybersecurity

Into one cohesive research direction.

Closing Reflection

This project reshaped how I view AI development.

The future is not AI replacing engineers.

The future is engineers who understand:

Statistical learning
Model interpretability
Cryptographic integrity
Adversarial robustness

Google Gemini did not build this system for me.

It sharpened my questions, accelerated my iterations, and forced clarity in my thinking.

The responsibility — methodological, ethical, and security-oriented — remained human.

And that is precisely how AI should be used.

Frist Challenge

nasihuyxbigmo — Sat, 28 Feb 2026 13:16:39 +0000

The Community
MindCare Community was built for the Indonesian community particularly young adults, students, and general public — who often lack easy access to mental health screening tools and psychological support. Mental health literacy is still growing in Indonesia, and many people don't know where to start. This platform is designed as a preventive, educational, and supportive first step before professional intervention.

What I Built
MindCare Community — a full-stack mental health platform built with Next.js 15, TailwindCSS, and Groq AI (LLaMA 3).

Key features:

🧠 Mental Health Screening — PHQ-9 and SRQ-20 standardized instruments with risk scoring and PDF reports
🫀 Physical Health Screening — Cardiovascular and metabolic risk assessment
📊 Daily Mood Tracker — Trend analysis with self-care recommendations
💬 AI Psychological Consultation — Emotion detection, 4 AI personas (Supportive Friend, Calm Therapist, Direct Coach, Spiritual Guide), Calm Mode UI, and Emotional Energy Battery indicator
🎙️ Real-Time Voice AI Mode — Speak directly to AI: mic → speech-to-text → AI → text-to-speech
🧩 MBTI Personality Test — 16-type assessment with cognitive function-grounded AI profiles and shareable result images
🌙 Glassmorphism UI — Light/dark mode persisted in localStorage, floating gradient orbs, smooth transitions
Demo
🔗 Live: https://preventive-health-screening-tool-gt.vercel.app/

Code

prasbara / Preventive-Health-Screening-Tool

Health Risk Self-Assessment Website

A comprehensive Next.js 14+ web application for health risk self-assessment with AI psychological consultation featuring real-time facial emotion detection.

🌟 Features

Health Assessment

Interactive health risk assessment form
Auto-calculating BMI
Rule-based risk scoring (0-100)
Personalized lifestyle recommendations
Visual charts and analytics

AI Psychological Consultation

Real-time facial emotion detection (client-side only)
AI-powered supportive chat with emotion context
Crisis detection and safety protocols
Privacy-first: no video/images stored

Admin Dashboard

Secure login system
Assessment statistics and analytics
Risk distribution visualization
Chat session analytics
Emotion distribution tracking

🚀 Getting Started

Prerequisites

Node.js 18+
npm or yarn

Installation

Clone the repository:

git clone <your-repo-url>
cd "Kesehatan Tester"

Install dependencies:

npm install

Set up environment variables: Create a .env.local file in the root directory with:

OPENAI_API_KEY=your_openai_key
GROQ_API_KEY=your_groq_key
GROQ_API_KEY_FALLBACK_1=your_fallback_key_1
GROQ_API_KEY_FALLBACK_2=your_fallback_key_2
ADMIN_EMAIL=yesii@admin
ADMIN_PASSWORD=yesii123
SESSION_SECRET=your_session_secret

Download face-api.js…

View on GitHub

How I Built It
Stack:

Framework: Next.js 15 (App Router)
Styling: TailwindCSS with custom glassmorphism design system
AI: Groq API (LLaMA 3.3-70B) with 9 API key rotation for load balancing and zero-downtime failover
Database: NeonDB (PostgreSQL) via Prisma
Auth: Iron Session (JWT-based admin sessions)
Speech: Web Speech API (SpeechRecognition + speechSynthesis)
Deployment: Vercel
Architecture highlights:

Dynamic multi-layer system prompt: persona × MBTI × calm mode × detected emotion × energy state — all composited per request
MBTI profiles grounded in cognitive function theory (Ni/Ne/Si/Se/Ti/Te/Fi/Fe), not stereotypes
Voice engine with silence detection, emotion heuristics, and crisis safety guard
Glassmorphism CSS design system using CSS custom properties with dark/light mode class strategy and FOUC prevention