Browse categories
Explore
Fiverr Pro
English
$
USD
I will build custom federated unlearning, machine unlearning expert
Usman Khan
About this gig
I will implement DynFRU a custom Certified Federated Machine Unlearning system that securely removes a client's data from the global model while preserving (or even improving) accuracy.
Using an adaptive dynamic controller with Gradient Ascent, Adaptive Scrub, and Fisher-scaled Noise, I deliver utility-positive unlearning with strong backdoor resistance and certified forgetting guarantees.
Perfect for researchers and teams needing privacy-compliant federated learning solutions.
Get to know Usman Khan
Usman Khan
Usman Khan
- FromPakistan
- Member sinceAug 2020
- Avg. response time1 hour
Languages
Urdu, Pashto, English
I am an AI Research & Development Specialist dedicated to building high-performance, privacy-first AI. My expertise spans the full spectrum of Machine Learning and Deep Learning, from custom neural networks to scalable predictive models. I specialize in Federated Learning (FL) for decentralized data and the cutting-edge field of Federated Unlearning (FU), implementing advanced frameworks to ensure data privacy and regulatory compliance. Whether you need a robust MVP or the implementation of complex research, I provide optimized, documented code to turn your data into intelligence.
FAQ
What is Federated Unlearning?
Federated Unlearning is the process of removing a specific client’s data influence from a trained federated model without retraining the entire system from scratch. It helps comply with privacy laws like the "Right to be Forgotten."
What is DynFRU and why is it better?
DynFRU is my custom framework for Dynamic Fisher-Risk Controlled Certified Federated Unlearning. It uses an intelligent neural controller that adaptively balances Gradient Ascent, Adaptive Scrub, and Fisher-scaled Noise. This results in near-zero or positive utility (accuracy often stays the same.
Do you support backdoor poisoning attacks?
Yes. I specialize in unlearning under backdoor poisoning scenarios. I can test and show results for both normal clients and the malicious client.
Will the model accuracy drop after unlearning?
In most cases, the accuracy drop is very small (less than 0.3%). In several experiments, I achieved utility-positive results where accuracy actually increased after unlearning.
What kind of models do you work with?
I primarily use a heterogeneous deep ensemble. I can adapt the solution to other models or your custom architecture if needed.
Do you provide the full code and explanations?
Yes. You will receive clean, well-documented Python code, full training + unlearning pipeline, evaluation metrics, visualizations, and detailed explanations.
Can you customize the solution for my dataset?
Absolutely. I can adjust the number of clients, poisoning level, unlearning strength, and other parameters according to your specific requirements.
What metrics do you provide?
I deliver: Global Accuracy, MIA AUC, Forget Quality, Backdoor ASR, AUS, Dynamic Forgetting Efficiency (DFE), and Certified Forgetting Bound (ε).
How long does it take to complete the gig?
Most standard implementations take 3–7 days, depending on complexity and custom requirements. I will give you a clear timeline after discussing your needs.
Do you offer revisions?
Yes, I offer unlimited revisions until you are fully satisfied with the results.
