IncludeHelp_logo
Skip Navigation Links

Home » Trending Technologies MCQs

Federated Learning MCQs (Multiple-Choice Questions)

Practice Federated Learning MCQs to test your knowledge of decentralized machine learning, client-server training, Federated Averaging, non-IID data, model aggregation, communication efficiency, privacy, secure aggregation, and federated optimization. These questions are useful for students, machine learning engineers, data scientists, AI developers, and professionals working with privacy-preserving distributed AI systems. The set includes both foundational and practical questions covering modern Federated Learning systems.

Federated Learning MCQs

These Federated Learning multiple-choice questions cover important concepts such as federated clients, servers, local training, Federated Averaging, client selection, local epochs, weighted aggregation, non-IID data, statistical heterogeneity, system heterogeneity, communication constraints, model-update compression, differential privacy, secure aggregation, poisoning attacks, backdoor attacks, personalization, cross-device learning, cross-silo learning, federated analytics, and federated evaluation. This set combines conceptual, technical, and scenario-based questions to help test your understanding of Federated Learning systems.

Federated Learning MCQs cover the technologies and methods used to train machine-learning models across decentralized datasets without requiring all raw training data to be centrally collected. Each question includes an answer and explanation.

List of Federated Learning MCQs

The following Federated Learning multiple-choice questions cover FedAvg, client participation, distributed optimization, privacy mechanisms, communication efficiency, heterogeneous data, security threats, personalization, and practical deployment scenarios.

1. What is the primary objective of Federated Learning?

  1. Move all training data to a central server
  2. Train a shared machine-learning model while keeping training data distributed across clients
  3. Eliminate the need for machine-learning models
  4. Store all client data in a public database

Answer: B) Train a shared machine-learning model while keeping training data distributed across clients

Explanation:

Federated Learning enables multiple clients to collaboratively train a shared model while keeping their local training data at the participating devices or organizations.

2. In a typical centralized Federated Learning architecture, where is the training data located?

  1. Only on the central server
  2. Across participating clients
  3. Only in a public cloud database
  4. Only inside the model parameters

Answer: B) Across participating clients

Explanation:

The defining characteristic of Federated Learning is that training data remains distributed across clients instead of being consolidated into a central training repository.

3. What is the role of the central server in a common Federated Learning architecture?

  1. Collect and store every client's raw training data
  2. Coordinate training rounds and aggregate client model updates
  3. Generate all client training examples
  4. Replace every client optimizer

Answer: B) Coordinate training rounds and aggregate client model updates

Explanation:

The server commonly distributes the current global model, coordinates client participation, receives client updates, and produces an updated global model through aggregation.

4. What is Federated Averaging (FedAvg)?

  1. An algorithm that aggregates locally trained model updates to produce a new global model
  2. A method for encrypting model parameters
  3. A database replication protocol
  4. A neural-network activation function

Answer: A) An algorithm that aggregates locally trained model updates to produce a new global model

Explanation:

FedAvg is a foundational Federated Learning algorithm in which selected clients perform local training and the server aggregates their resulting model updates, commonly using a data-weighted average.

5. Which sequence best represents a typical FedAvg training round?

  1. Server selects clients, sends the model, clients train locally, updates are aggregated, and the global model is updated
  2. Clients upload raw data, server trains, and clients delete their models
  3. Server permanently trains a separate model for each client
  4. Clients exchange their raw datasets directly

Answer: A) Server selects clients, sends the model, clients train locally, updates are aggregated, and the global model is updated

Explanation:

A typical round starts with client selection and model distribution, followed by local computation and client-to-server update transmission. The server then aggregates the updates into the next global model.

6. What does a client typically receive at the beginning of a federated training round?

  1. The current global model parameters
  2. Every other client's raw dataset
  3. The server's private training data
  4. The complete history of every client's examples

Answer: A) The current global model parameters

Explanation:

In standard centralized FL, the server sends the current global model or relevant model parameters to selected clients so they can perform local training.

7. What does a client normally do with the global model before sending an update?

  1. Train or fine-tune it using its local data
  2. Publish it publicly
  3. Delete all model parameters
  4. Replace it with raw training records

Answer: A) Train or fine-tune it using its local data

Explanation:

The client uses its locally stored data to perform one or more local optimization steps on the received global model before sending an update or resulting model information to the server.

8. Why is communication efficiency particularly important in Federated Learning?

  1. Clients may have limited bandwidth, intermittent connectivity, or expensive communication
  2. Federated models never require communication
  3. All clients always operate inside the same data center
  4. Communication has no effect on training time

Answer: A) Clients may have limited bandwidth, intermittent connectivity, or expensive communication

Explanation:

Many FL deployments involve mobile or edge devices with constrained networks. Reducing communication rounds and update sizes can therefore have a major impact on system performance.

9. What is a federated training round?

  1. A coordinated cycle in which selected clients receive a model, perform local computation, and contribute updates
  2. A single prediction made by one client
  3. A database backup operation
  4. A hardware reboot

Answer: A) A coordinated cycle in which selected clients receive a model, perform local computation, and contribute updates

Explanation:

A training round represents one coordinated iteration of federated computation involving a selected group of clients and the server.

10. Why are only a subset of clients commonly selected in a cross-device FL round?

  1. It is often impractical to coordinate the entire population of potentially available devices at once
  2. Only one client can ever participate in Federated Learning
  3. Clients are required to share their passwords
  4. All clients must be physically located together

Answer: A) It is often impractical to coordinate the entire population of potentially available devices at once

Explanation:

Cross-device FL may involve very large client populations. Sampling a manageable subset per round reduces coordination and communication requirements.

11. What does non-IID data mean in Federated Learning?

  1. Client datasets may have different statistical distributions
  2. All clients have identical datasets
  3. Clients have no training data
  4. Every client uses a different programming language

Answer: A) Client datasets may have different statistical distributions

Explanation:

Federated datasets are often non-IID, meaning the distribution of features, labels, sample counts, or other characteristics can vary substantially between clients.

12. Why can non-IID data make FedAvg more difficult to optimize?

  1. Local client updates may point in different optimization directions
  2. Clients cannot perform local training
  3. Model parameters become impossible to represent
  4. Communication automatically becomes free

Answer: A) Local client updates may point in different optimization directions

Explanation:

When clients have different data distributions, their local objectives can differ. Their updates may therefore be inconsistent or biased relative to the global objective.

13. What is statistical heterogeneity in Federated Learning?

  1. Variation in data distributions, quantities, or characteristics across clients
  2. Variation in CPU clock speeds only
  3. Variation in network cables only
  4. Variation in programming languages only

Answer: A) Variation in data distributions, quantities, or characteristics across clients

Explanation:

Statistical heterogeneity refers to differences in the data held by clients, including different label distributions, feature distributions, sample sizes, and other characteristics.

14. What is system heterogeneity in Federated Learning?

  1. Differences in client hardware, network connectivity, availability, storage, or computational resources
  2. Differences only in label distributions
  3. Differences in database schemas only
  4. Differences in encryption algorithms only

Answer: A) Differences in client hardware, network connectivity, availability, storage, or computational resources

Explanation:

System heterogeneity occurs because participating devices can differ significantly in their computational capabilities, battery state, network conditions, and availability.

15. What is client drift in federated optimization?

  1. The divergence of local model updates caused by clients optimizing against different local data distributions
  2. The physical movement of mobile devices
  3. The loss of network connectivity only
  4. The deletion of client datasets

Answer: A) The divergence of local model updates caused by clients optimizing against different local data distributions

Explanation:

With heterogeneous local objectives, clients can move their models in different directions during local training. This divergence is commonly referred to as client drift.

16. What is the effect of increasing the number of local epochs in FedAvg?

  1. It can reduce communication frequency but may increase local computation and client drift
  2. It always eliminates non-IID problems
  3. It guarantees higher global accuracy
  4. It removes the need for server aggregation

Answer: A) It can reduce communication frequency but may increase local computation and client drift

Explanation:

More local training can make each communication round perform more computation, potentially reducing the number of rounds required. However, excessive local training can worsen divergence between client updates on heterogeneous data.

17. In weighted FedAvg, why might clients contribute different weights to the global update?

  1. Clients may have different numbers of local training examples
  2. Clients use different operating systems
  3. Clients have different screen sizes
  4. Clients have different IP addresses

Answer: A) Clients may have different numbers of local training examples

Explanation:

A common weighted FedAvg formulation gives client updates weights proportional to the amount of local training data, so clients with more examples can have greater influence on the aggregated model.

18. What is a major advantage of keeping raw training data on clients?

  1. It reduces the need to centrally collect raw data
  2. It guarantees that no information can leak through model updates
  3. It eliminates all security risks
  4. It guarantees model fairness

Answer: A) It reduces the need to centrally collect raw data

Explanation:

Keeping data local can reduce centralized data exposure and support privacy objectives. However, model updates and system behavior can still create privacy and security risks.

19. Does Federated Learning by itself guarantee complete privacy?

  1. No, model updates can potentially leak information about local data
  2. Yes, because raw data never leaves the client
  3. Yes, because all FL models are encrypted
  4. No, because FL requires raw data to be uploaded

Answer: A) No, model updates can potentially leak information about local data

Explanation:

Keeping raw data local is not equivalent to a formal privacy guarantee. Updates can potentially expose information, so additional mechanisms such as secure aggregation or differential privacy may be required.

20. What is Secure Aggregation designed to protect in Federated Learning?

  1. Individual client updates from being directly inspected by the server during aggregation
  2. The physical location of every client
  3. The accuracy of the global model
  4. The model architecture from all clients

Answer: A) Individual client updates from being directly inspected by the server during aggregation

Explanation:

Secure aggregation allows a server to obtain an aggregate of client contributions without directly observing each individual contribution in unaggregated form.

21. What is the main purpose of differential privacy in Federated Learning?

  1. Limit the information that model training or released results reveal about individual data contributions
  2. Increase the number of clients automatically
  3. Replace the model optimizer
  4. Guarantee perfect model accuracy

Answer: A) Limit the information that model training or released results reveal about individual data contributions

Explanation:

Differential privacy provides a formal framework for bounding privacy loss. In federated training, clipping and noise can be applied to client contributions to provide user-level privacy guarantees under a defined accounting framework.

22. Why is client update clipping useful when applying user-level differential privacy?

  1. It bounds the maximum influence of an individual client's update before noise is added
  2. It guarantees that every client has identical data
  3. It removes the need for aggregation
  4. It increases the size of every update

Answer: A) It bounds the maximum influence of an individual client's update before noise is added

Explanation:

Clipping limits the norm of individual client updates. This bounds their sensitivity and allows calibrated noise to provide differential privacy.

23. What is a common trade-off when adding differential privacy noise to federated updates?

  1. Stronger privacy can reduce model utility or accuracy
  2. Privacy noise always increases accuracy
  3. Noise eliminates all communication
  4. Noise makes all client datasets IID

Answer: A) Stronger privacy can reduce model utility or accuracy

Explanation:

Adding noise protects privacy but can make the optimization process less precise. The resulting privacy-utility trade-off must be evaluated for the particular model and workload.

24. What is a poisoning attack in Federated Learning?

  1. An attack in which malicious clients manipulate training data or model updates to influence the global model
  2. An attack that only steals server hardware
  3. A method for compressing model updates
  4. A technique for increasing client participation

Answer: A) An attack in which malicious clients manipulate training data or model updates to influence the global model

Explanation:

Poisoning attacks attempt to compromise the learning process by manipulating local data or submitted updates, potentially degrading model performance or changing model behavior.

25. What is a backdoor attack in Federated Learning?

  1. An attack that attempts to make the model behave maliciously for selected trigger inputs while maintaining normal behavior on many other inputs
  2. An attack against physical server doors
  3. A communication compression method
  4. A privacy accounting technique

Answer: A) An attack that attempts to make the model behave maliciously for selected trigger inputs while maintaining normal behavior on many other inputs

Explanation:

Backdoor attacks introduce targeted malicious behavior into the model. The global model may continue performing normally on ordinary validation data, making detection more difficult.

26. Why can simple averaging be vulnerable to malicious client updates?

  1. A malicious client can submit an unusually manipulated update that influences the aggregate
  2. Averaging automatically detects every attack
  3. All clients are guaranteed to be honest
  4. Simple averaging encrypts every update

Answer: A) A malicious client can submit an unusually manipulated update that influences the aggregate

Explanation:

Basic averaging does not inherently distinguish honest updates from malicious ones. Robust aggregation methods can be used to reduce the influence of anomalous or adversarial contributions.

27. What is robust aggregation in Federated Learning intended to address?

  1. The influence of anomalous or potentially malicious client updates
  2. Only network latency
  3. Only dataset storage
  4. Only model serialization

Answer: A) The influence of anomalous or potentially malicious client updates

Explanation:

Robust aggregation methods attempt to reduce the impact of abnormal client contributions, helping defend the training process against certain poisoning and Byzantine-style behaviors.

28. What is cross-device Federated Learning?

  1. Federated Learning involving potentially very large numbers of consumer or edge devices
  2. Federated Learning limited to two data centers
  3. A centralized training method using one server
  4. A technique for database synchronization only

Answer: A) Federated Learning involving potentially very large numbers of consumer or edge devices

Explanation:

Cross-device FL commonly targets large populations of devices such as smartphones or IoT devices, where availability and network conditions can vary substantially.

29. What is cross-silo Federated Learning?

  1. Federated Learning involving a relatively small number of organizations or data silos
  2. Federated Learning involving billions of mobile devices only
  3. Centralized training on one dataset
  4. Model training without any communication

Answer: A) Federated Learning involving a relatively small number of organizations or data silos

Explanation:

Cross-silo FL is commonly used when organizations such as hospitals, banks, or enterprises collaboratively train models while keeping their institutional data within their own environments.

30. Which characteristic is particularly common in cross-device FL?

  1. Unreliable client availability and heterogeneous device resources
  2. All clients having identical hardware
  3. Unlimited network bandwidth
  4. Permanent participation of every client

Answer: A) Unreliable client availability and heterogeneous device resources

Explanation:

Mobile and edge devices can be offline, battery-constrained, bandwidth-limited, or computationally heterogeneous. FL systems therefore commonly select clients based on availability and system conditions.

31. What is client dropout?

  1. The failure of a participating client to complete or submit its expected computation
  2. The deletion of the global model
  3. The removal of the server optimizer
  4. The conversion of raw data to synthetic data

Answer: A) The failure of a participating client to complete or submit its expected computation

Explanation:

Clients may disconnect, run out of battery, lose network access, or otherwise fail to finish a round. Federated protocols need to tolerate such partial participation.

32. Why is partial client participation important in Federated Learning?

  1. Real-world clients may be unavailable, and coordinating every client can be impractical
  2. Every client must participate for every round
  3. It eliminates non-IID data
  4. It guarantees privacy

Answer: A) Real-world clients may be unavailable, and coordinating every client can be impractical

Explanation:

Federated systems are often designed around a subset of clients participating in each round. The global training algorithm must therefore work despite changing participant sets.

33. What is model-update compression used for in Federated Learning?

  1. Reducing the amount of information that must be transmitted between clients and the server
  2. Increasing the number of raw training records
  3. Removing all model parameters
  4. Replacing local training with centralized training

Answer: A) Reducing the amount of information that must be transmitted between clients and the server

Explanation:

Quantization, sparsification, structured updates, and other compression approaches can reduce communication costs by decreasing the size of transmitted model information.

34. What is quantization in the context of federated model updates?

  1. Representing numerical model information with fewer bits or a smaller numerical representation
  2. Adding more training examples to each client
  3. Encrypting every model parameter with a different key
  4. Deleting the model architecture

Answer: A) Representing numerical model information with fewer bits or a smaller numerical representation

Explanation:

Quantization reduces the precision used to represent model updates. This can lower communication costs but may introduce approximation error.

35. What is sparsification of federated model updates?

  1. Transmitting only a selected subset of update values instead of the complete dense update
  2. Increasing every model value to maximum precision
  3. Duplicating every client update
  4. Replacing model updates with raw training examples

Answer: A) Transmitting only a selected subset of update values instead of the complete dense update

Explanation:

Sparsification reduces communication by transmitting selected model-update components, such as values with large magnitudes, while handling omitted information according to the chosen algorithm.

36. What is personalized Federated Learning?

  1. Federated Learning approaches that adapt the learned model to individual clients or client groups
  2. A method that forces every client to use exactly the same final model
  3. A technique for encrypting personalized data
  4. A centralized recommendation system

Answer: A) Federated Learning approaches that adapt the learned model to individual clients or client groups

Explanation:

Personalized FL recognizes that different clients may have different data distributions or objectives and therefore may benefit from models adapted to their local characteristics.

37. Why can a single global model perform poorly for some clients?

  1. Client data distributions may differ substantially from the aggregate distribution
  2. The global model never receives client updates
  3. All client datasets are guaranteed to be identical
  4. Federated models cannot perform inference

Answer: A) Client data distributions may differ substantially from the aggregate distribution

Explanation:

Strong statistical heterogeneity can cause a globally optimized model to perform unevenly across clients. Personalization methods can address this by adapting models to local distributions.

38. What is federated analytics?

  1. Computing aggregate statistics across decentralized client data without centrally collecting all raw data
  2. Training only one neural network on a central dataset
  3. Analyzing server CPU metrics only
  4. Compressing model parameters

Answer: A) Computing aggregate statistics across decentralized client data without centrally collecting all raw data

Explanation:

Federated analytics applies federated computation concepts to statistics and other aggregate analyses rather than necessarily training a machine-learning model.

39. Which scenario is a good example of federated analytics?

  1. Computing the frequency of a feature across many devices without collecting every user's raw record centrally
  2. Copying every device database to a server
  3. Training a model on one centralized dataset
  4. Compressing a mobile application

Answer: A) Computing the frequency of a feature across many devices without collecting every user's raw record centrally

Explanation:

Federated analytics can calculate aggregate statistics across distributed datasets while avoiding the need to centralize all underlying records.

40. Which metric is particularly useful for evaluating whether a federated model performs consistently across clients?

  1. Per-client evaluation metrics
  2. Server disk capacity only
  3. Number of network cables
  4. Model file name

Answer: A) Per-client evaluation metrics

Explanation:

Aggregate accuracy can hide large differences between clients. Evaluating performance across clients can reveal disparities caused by heterogeneous data distributions.

41. A smartphone keyboard model is trained using user typing data that should remain on individual devices. Which architecture is most appropriate?

  1. Cross-device Federated Learning
  2. Centralized training with raw keystroke uploads
  3. Public data publication
  4. Manual model training on one device

Answer: A) Cross-device Federated Learning

Explanation:

A smartphone keyboard is a typical cross-device FL scenario because training data can remain on devices while locally computed updates contribute to a shared model.

42. Several hospitals want to train a common disease-prediction model while keeping patient records within their own institutions. Which approach best fits this requirement?

  1. Cross-silo Federated Learning
  2. Centralized raw-data pooling
  3. Public database publication
  4. Single-client training

Answer: A) Cross-silo Federated Learning

Explanation:

Hospitals represent organizational data silos. Cross-silo FL can allow them to collaborate on model training while maintaining local control over their patient datasets.

43. A federated model is performing poorly because each client has a highly different label distribution. Which problem is most directly indicated?

  1. Statistical heterogeneity
  2. Network encryption
  3. Model serialization
  4. Hardware virtualization

Answer: A) Statistical heterogeneity

Explanation:

Different label distributions across clients are a form of statistical heterogeneity and can make the global optimization problem more challenging.

44. A federated training system has extremely large model updates and expensive mobile network connections. Which optimization is most directly relevant?

  1. Model-update compression or sparsification
  2. Increasing raw dataset size
  3. Publishing all training data
  4. Adding more model layers without considering communication

Answer: A) Model-update compression or sparsification

Explanation:

Compression and sparsification can reduce the amount of data transmitted during federated rounds, which is particularly valuable for bandwidth-constrained clients.

45. A malicious client repeatedly submits manipulated model updates intended to make a cancer classifier misclassify a specific trigger pattern. What type of attack is this?

  1. Backdoor attack
  2. Model compression
  3. Client dropout
  4. Federated evaluation

Answer: A) Backdoor attack

Explanation:

A backdoor attack attempts to implant targeted behavior that is activated by a particular trigger while allowing the model to behave normally on many standard inputs.

46. A federated learning system uses secure aggregation but does not use differential privacy. Which statement is most accurate?

  1. Secure aggregation can hide individual updates from the server during aggregation, but it is not equivalent to differential privacy
  2. Secure aggregation guarantees formal differential privacy
  3. Differential privacy is impossible with secure aggregation
  4. Secure aggregation eliminates all privacy risks

Answer: A) Secure aggregation can hide individual updates from the server during aggregation, but it is not equivalent to differential privacy

Explanation:

Secure aggregation and differential privacy address different privacy concerns. Secure aggregation can prevent direct inspection of individual contributions, while differential privacy provides a mathematical bound on information leakage under defined assumptions.

47. A federated model has good overall accuracy but performs poorly for a small group of clients with uncommon data distributions. What is the most appropriate investigation?

  1. Analyze per-client performance and consider personalization or methods designed for heterogeneous data
  2. Ignore the affected clients because aggregate accuracy is high
  3. Delete the clients' data
  4. Increase communication without analyzing the data distribution

Answer: A) Analyze per-client performance and consider personalization or methods designed for heterogeneous data

Explanation:

Aggregate metrics can hide poor performance for specific client groups. Per-client evaluation can reveal distribution-related problems and help determine whether personalization or another federated optimization strategy is appropriate.

48. A server selects 100 clients for a round, but only 70 complete local training and return updates. What should a robust federated system generally do?

  1. Handle partial participation and aggregate the valid updates according to the protocol
  2. Upload the missing clients' raw data
  3. Assume the missing clients trained successfully
  4. Discard the entire global model permanently

Answer: A) Handle partial participation and aggregate the valid updates according to the protocol

Explanation:

Client availability can change during a federated round. Practical FL protocols are designed to tolerate client dropout and proceed using the contributions that successfully satisfy the aggregation requirements.

49. A privacy-sensitive federated system wants to reduce the server's ability to inspect individual client updates while also limiting information leakage from the final model. Which combination is most appropriate to evaluate?

  1. Secure aggregation together with differential privacy
  2. Only model compression
  3. Only client-side caching
  4. Only increasing the number of local epochs

Answer: A) Secure aggregation together with differential privacy

Explanation:

Secure aggregation can protect individual contributions from direct server inspection, while differential privacy can limit information leakage through model outputs or aggregated updates under a defined privacy mechanism and accounting framework.

50. A company is deploying Federated Learning across millions of mobile devices. The devices have non-IID data, intermittent connectivity, limited bandwidth, and different hardware capabilities. The company also wants strong privacy protection. Which architecture is the most technically appropriate?

  1. Centralize all raw device data and train one model on the server
  2. Use client sampling, local training, communication-efficient aggregation, secure aggregation, differential privacy where required, and evaluation that accounts for client heterogeneity
  3. Require every device to participate in every training round and upload its complete dataset
  4. Use simple averaging without monitoring client behavior or privacy risks

Answer: B) Use client sampling, local training, communication-efficient aggregation, secure aggregation, differential privacy where required, and evaluation that accounts for client heterogeneity

Explanation:

A large cross-device deployment must account for partial participation, non-IID data, unreliable connectivity, and resource constraints. Communication-efficient training can reduce network costs, while secure aggregation and differential privacy can address different privacy risks. Per-client evaluation is also important because aggregate metrics may hide poor performance on specific client populations.

Comments and Discussions!

Load comments ↻



Top MCQs
Top Programs/Examples
Top Tutorials
About
Student's Section
Subscribe

Copyright © 2026 www.includehelp.com. All rights reserved.