Federated Learning: Privacy-Preserving Edge AI Fundamentals Quiz Quiz

Explore the essentials of federated learning and its privacy-preserving techniques for Edge AI. Assess your understanding of distributed model training, data privacy, security concerns, and the role of local devices in decentralized machine learning systems.

1. Definition of Federated Learning

   Which statement best describes federated learning in the context of Edge AI?

   1. It requires devices to work independently, never sharing any information.
   2. It trains AI models by aggregating updates from multiple local devices without collecting raw data.
   3. It transfers all raw data to a central server for model training.
   4. It uses one device at a time to train the model sequentially.

   Explanation: Federated learning aggregates model updates from devices without gathering raw data, preserving user privacy. Sending raw data to a central server, as stated in another option, contradicts privacy goals. Devices do share some information, unlike the option suggesting no sharing. Training on one device at a time sequentially is inefficient and not how federated learning works.

2. Privacy Benefit of Federated Learning

   In federated learning, what main privacy benefit is provided to users compared to traditional centralized training?

   1. User data remains on local devices instead of being sent to a central server.
   2. User data is broadcasted to all participating devices.
   3. User data is deleted before training begins.
   4. All personal data is encrypted and sent to the cloud.

   Explanation: Keeping user data on local devices reduces privacy risks by minimizing exposure. Encrypting and sending data to the cloud still involves transmission and possible exposure. Deleting user data before training would prevent any learning. Broadcasting user data conflicts with privacy preservation.

3. Role of the Edge Device

   What is the primary function of an edge device in the federated learning process?

   1. Running all possible models independently and never interacting with others.
   2. Collecting data from nearby devices for central training.
   3. Locally updating the AI model using its own data and sharing updates, not raw data.
   4. Broadcasting all raw data to a server for aggregation.

   Explanation: Edge devices perform local model training using their data and share only improved model updates, preserving privacy. Running models independently without sharing updates impedes learning. Collecting data from other devices undermines decentralization. Broadcasting raw data defeats the privacy intent.

4. Model Aggregation Techniques

   Which method is commonly used in federated learning to combine updates from multiple devices?

   1. Randomly selecting one device's model as the new global model.
   2. Sum up all raw data from devices and retrain globally.
   3. Using only the largest update for global aggregation.
   4. Averaging the weights from all participating local models.

   Explanation: Averaging model weights is a standard approach to aggregate distributed model updates. Random selection fails to use all learned knowledge. Using only the largest update ignores most devices' contributions. Summing raw data and re-training violates privacy.

5. Data Heterogeneity Challenge

   How does federated learning address the challenge of different data distributions (heterogeneity) on edge devices?

   1. It ignores differences in data and treats all updates equally.
   2. It transfers all data to the cloud to standardize the dataset.
   3. It forces all devices to use identical datasets before training.
   4. It aggregates diverse model updates, allowing the global model to generalize across varied data.

   Explanation: By collecting diverse updates, federated learning helps the global model adapt to different data conditions. Forcing identical datasets is impractical and not privacy-preserving. Ignoring data differences reduces effectiveness, while transferring data to the cloud undermines privacy.

6. Communication Round Concept

   In federated learning, what does a 'communication round' refer to?

   1. A session where devices send all their data to each other.
   2. A period when the server asks devices for their passwords.
   3. A cycle where devices train locally and send their updates to a central server.
   4. The continuous streaming of raw data from all devices.

   Explanation: A communication round involves local training followed by the transfer of model updates, not raw data. Devices do not exchange all their data with each other, nor do they continuously stream data. Server password requests are unrelated to federated learning.

7. Security Threats

   Which security threat may affect federated learning during the transmission of model updates?

   1. Eavesdropping on model updates sent between devices and the server.
   2. Stealing raw data directly from all devices.
   3. Sending encrypted updates using secure communication protocols.
   4. Sharing updates only on paper rather than electronically.

   Explanation: Eavesdropping intercepts model updates and poses a risk, even if data is not sent directly. Stealing raw data is less likely since it is not transmitted. Sharing updates on paper is unrealistic and irrelevant. Sending encrypted updates is actually a mitigation strategy, not a threat.

8. Differential Privacy Use

   How can differential privacy be applied in federated learning?

   1. Broadcasting all updates unmodified to every participant.
   2. Deleting all updates after each training round.
   3. Sending only complete datasets to the central server.
   4. Adding random noise to model updates before sharing them with the server.

   Explanation: Adding noise helps protect individual data contributions, enhancing privacy. Broadcasting unmodified updates reduces privacy. Deleting all updates would impede learning, while sending complete datasets contradicts the privacy-preserving goals.

9. Example Application

   Which of the following is an example where federated learning is beneficial?

   1. Sharing user passwords for collaborative model updates.
   2. Requiring only one device to perform all the training with synthetic data.
   3. Storing all user data on a central server for analysis.
   4. Training a predictive text model using user data stored locally on personal devices.

   Explanation: Federated learning allows models like predictive text to learn from user data without transferring it off the device. Storing all data centrally undermines privacy. Training on just one device is less effective and not federated. Sharing passwords is unsafe and irrelevant.

10. Edge AI Resource Requirements

    What is an important requirement for edge devices participating in federated learning?

    1. No storage capacity required on the device.
    2. Sufficient computing resources to locally train the AI model.
    3. Complete collection of data from other devices.
    4. Permanent connection to the server at all times.

    Explanation: Edge devices need enough computing power to update models with local data. Permanent connection is unnecessary as updates can be sent intermittently. Collecting data from others contradicts federated learning's decentralized design. Some storage is essential for model and temporary data.