not approved
Decentralized Next-Gen Edge AI
Current Project Status
Unfunded
Amount
Received
$0
Amount
Requested
$50,000
Percentage
Received
0.00%
Solution

A decentralized framework is key to build next-gen edge AI that enables model efficacy while making it safer and secure - both user data and model parameters are only accessible to the end-user.

Problem

AI has proven its adeptness in real-world scenarios but is exploitatively dangerous at the hands of big tech giants. Centralization is a big threat to user-centric, privacy-preserving next-gen edge AI

Impact / Alignment
Feasibility
Auditability

Team

2 members

Decentralized Next-Gen Edge AI

Please describe your proposed solution.

This proposal is a continuation of our independent research work, in which we created a proof-of-concept decentralized framework to train and deploy edge AI models on the Ethereum blockchain. Our project will be integrated with Cardano for faster, cheaper, and more secure transactions.

The Big Problem

Interactive Web 2.0 is progressively exploiting human decision-making and is unprecedently influencing behavior via centralized AI algorithms. Such large-scale AI systems ought to be built and evaluated to perform a task on a public distributed ledger platform, similar to how primates and humans have successfully evolved higher cognitive intelligence within social constructs. Yet, following in the footsteps of big tech research, AI benchmarks and algorithms rely on centralized datasets and algorithms that pose a threat to secure closed-loop behavior and learning outcomes, both commonly modulated in biological organisms via social interactions.

The AI Conundrum

State-of-the-art AI systems, such as computer vision algorithms, rely heavily on deep learning algorithms trained on large datasets. The most significant strides in computer vision and deep neural networks were spurred by the rise of data-driven systems, leading to some truly astonishing capabilities, from the ability to achieve human-like (and even super-human) levels of performance under ideal viewing conditions on certain vision tasks to the unsettling ability to realistically replace faces and people in high-definition video. However, such cutting-edge data-driven systems require unprecedentedly large datasets and are unlikely to scale with increasing task complexity. The corresponding networks ingesting this data have grown vast in size and scale. Large datasets become difficult to distribute and test against and even more difficult to collect. Only a handful of organizations possess the resources required to collect and generate the cutting-edge datasets used at the forefront of deep learning. Since the volume of data and computational power is often insufficient to train locally, central servers have enabled researchers to train ever-larger networks, optimizing and pushing the limits of deep learning models. These centralized cloud solutions have inherent disadvantages, such as increased data traffic, potential loss of confidentiality, privacy, and security of user data. Consequently, federated learning proposed by Google addresses some of these challenges posed by central learning. In federated learning, the training of AI models is done locally and the model parameters are handled by a central server. This, however, intrinsically makes the entire system vulnerable to a model inversion attack and a single point of failure that halts the federated learning process.

Coming-of-age of Blockchain Technology:

Before the maturing of blockchain platforms, the idea of integrating them with machine learning was limited to marketplaces. Such systems stored already trained models in smart contracts for competitions and did not allow for continual updating and collaborative training. The blockchain smart contract enables model evolution and storage via IPFS, which is typically handled by a central server in federated learning. Without a central entity, blockchain-based federated learning is cryptographically secure while preserving the data privacy of each node.

OUR SOLUTION:

Our current work is implemented on the Ethereum Smart Contract Platform based on the architecture below:

Image File

Our framework seamlessly integrates state-of-the-art deep learning models and serves as a good benchmark for blockchain-AI developers. Please watch the Youtube videos to have more information about the literature and how our work is novel and different to existing work.

Image File

It has two modes of operations as explained on the Github page. The python script interacts with the blockchain and the IPFS without needing a Solidity interface. This work will be detailed in a publication at IEEE Blockchain and Beyond conference. The future work will address practical real-world considerations while porting to the Cardano blockchain - first via Milkomeda and next via native Plutus implementation.

Please describe how your proposed solution will address the Challenge that you have submitted it in.

Our Edge AI system on the Cardano Blockchain will be able to address two main challenges:-

  1. Proof of concept for a decentralized edge AI system on a blockchain. AI systems on a blockchain will prove to be more trustworthy than a centralized system.
  2. We believe our multi-chain, model agnostic framework will open doors for the use and validation of our architecture on actual edge devices and solve real-world problems such as biomedical image processing.

Our current solution on Ethereum incurs high gas fees and low throughput for the nodes of the Decentralized Federated Learning framework. Integrating with Cardano is the natural step for exploiting the advanced EUTXO capabilities besides cheaper and faster throughput.

Our framework guarantees secure, transparent, and fair onboarding of the nodes without the need for a central custodian. The smart contract allows parameter merging with equal rights for all nodes and protects machine learning models from corruption via an incentive mechanism. An immediate application of DFL is privacy-preserving federated machine learning in medicine with the added security of decentralization.

What are the main risks that could prevent you from delivering the project successfully and please explain how you will mitigate each risk?

Risk mitigation will be systematically carried out by creating our Plutus solution in a modular fashion. Thus we maintain code reusability and avoid total system failure due to unforeseen bugs.

One of the main challenges is to prevent model inversion attacks and we propose homomorphic encryption to be incorporated in this smart contract.

Another challenge is to design our Cardano-based framework to fully exploit the E-UTXO model. In particular, handling multiple UTXOs for the different edge nodes while updating the federated model via the smart contract. Our expertise as Plutus Pioneers will be handy and we have successfully handled similar issues in other Cardano projects.

Please provide a detailed plan, including timeline and key milestones for delivering your proposal.

Milestones

1 Month | Milestone 1 - Develop the Chain-agnostic Database System that stores ML models using IPFS (avoid storing on-chain limitations)- We need to port the Trained Models we made earlier and enable reading/writing chain agnostic.

2 Month | Milestone 2 - Develop the UI code to run the smart contract/node communication. Deploy a test network of 3-5 nodes and a server.

2 Month | Milestone 3 - Test the communication between the blockchain networks and the edge nodes. Local learning, no need to store training data on-chain (privacy). The actual training process takes place at the edge, thereby preserving the privacy and confidentiality of the users' data.

2 Month | Milestone 4 - Improved Incentive mechanism on Cardano. This is a critical piece of the ongoing work.

2 Months | Milestone 5 - Test the behavior of the federated learning model. Integration and testing of state-of-the-art models on sufficiently large data in a distributed fashion

Deliverables

Deliverable 1 - Source code of the Database management system ( IPFS Repository )

Deliverable 2 - Source code for the server/node communication. ( Github Repository )

Deliverable 3 - Source code for the decentralized federated learning framework on Cardano (Github Repository )

Our other proposals do not comprise the same team members and old work from previous funds has been completed.

Please provide a detailed budget breakdown.

Federated learning servers - $14K

Segregated Data Storage - $2K

Training scripts hosting servers - $5K

Web development 40 hours/week 8 weeks $25/hour - $8K

Smart contract development 40 hours/week 12 weeks $40/hour - $19K

User incentives - $2K

We aim to build a team of researchers, engineers and Plutus/Solidity developers within our company that specializes in tackling distributed AI on Blockchains for confidential deep learning. Our primary area of research includes computer vision and text data.

With a separate budget for incentivizing the pilot users, we will send updated models to our partner institutes that can help us improve the visibility of our solution, especially in Singapore and India. We aim to demonstrate the real use case of federated learning and improve individual nodes' performance. Our pilot partners will be research institutes within universities and other independent researchers too.

The public framework launch will be Q1 next year. Our pilot partners will have a demo before the end of this year.

Please provide details of the people who will work on the project.

We are an AI-Blockchain Solution Provider Company based in Singapore and India, primarily building on Cardano, with stakeholders in the National University of Singapore, Defense partners, and Biomedical companies that require confidential machine learning.

Dr. Bharath's Research Interests

  • Pattern recognition and computer vision
  • Event-based cameras for autonomous sensing and navigation
  • Object recognition and related areas such as scene understanding, face recognition, and object detection for silicon retinal, event-based cameras onboard unmanned aerial vehicles.
  • Control and Simulation, Image Classification using Invariant Features.

Publications

Bio for Sam:

● Plutus PBL 1st Cohort - Gimbalabs

● Founding Dev & Smart Contract Lead - Rarety.io

● Co-Founder - Fetachain.io

● 2021 Presidential Innovation Award - Government of India

● IIT Bombay & Mathworks Computational Agriculture Hackathon International Rank - 3.

Combining the academic and practical capacities of the co-founders, we are best placed to build a blockchain-based federated learning framework on the Cardano Blockchain that provides security, transparency, and governance to AI models.

If you are funded, will you return to Catalyst in a later round for further funding? Please explain why / why not.

Yes, to add more functionalities to what we have built from the funds we receive in this fund.

Please describe what you will measure to track your project's progress, and how will you measure these?

Our progress:

  • We will measure individual node performance as well as the overall federated learning model's performance. These two key metrics will show whether privacy-preserving learning has benefitted each node. At the ultimate level, the federated level model should be able to perform better than the individual nodes.
  • Commits to our open-source Github repo
  • Publishing documentation on our company website blog (work in progress)

KPIs:

  • Feedback from pilot users
  • Number of publicly available AI models generated by the framework for various AI tasks
  • Onboarding of new team members
  • DApp verification by IOG and auditing of smart contracts that realize AI model privacy.

What does success for this project look like?

Deployment of the decentralized federated learning framework using the Cardano blockchain. A full-scale AIDA system will be able to onboard real-world users with a specific model learning goal in a distributed fashion while preserving the confidentiality and privacy of the data. An example in our case is multiple hospitals being able to detect cancer cells more effectively using the federated model and avoid human errors. Our partners at the National University of Singapore, and the research institutes involved in this work, will be the pilot study members.

Please provide information on whether this proposal is a continuation of a previously funded project in Catalyst or an entirely new one.

NO

close

Playlist

  • EP2: epoch_length

    Authored by: Darlington Kofa

    3m 24s
    Darlington Kofa
  • EP1: 'd' parameter

    Authored by: Darlington Kofa

    4m 3s
    Darlington Kofa
  • EP3: key_deposit

    Authored by: Darlington Kofa

    3m 48s
    Darlington Kofa
  • EP4: epoch_no

    Authored by: Darlington Kofa

    2m 16s
    Darlington Kofa
  • EP5: max_block_size

    Authored by: Darlington Kofa

    3m 14s
    Darlington Kofa
  • EP6: pool_deposit

    Authored by: Darlington Kofa

    3m 19s
    Darlington Kofa
  • EP7: max_tx_size

    Authored by: Darlington Kofa

    4m 59s
    Darlington Kofa
0:00
/
~0:00