Privacy-Preserving On-Screen Activity Tracking and Classification in E-Learning Using Federated Learning. Shakas Technologies ( Galaxy of Knowledge) #11/A 2nd East Main Road, Gandhi Nagar, Vellore - 632006. Mobile : +91-9500218218 / 8220150373| land line- 0416- 3552723 Shakas Training & Development | Shakas Sales & Services | Shakas Educational Trust|IEEE projects | Research & Development | Journal Publication | Email : info@shakastech.com | shakastech@gmail.com | website: www.shakastech.com​ Facebook: https://www.facebook.com/pages/Shakas-Technologies

1. Base paper Title: Privacy-Preserving On-Screen Activity Tracking and Classification in E-
Learning Using Federated Learning
Modified Title: Federated Learning for Privacy-Preserving On-Screen Activity Tracking and
Classification in E-Learning
Abstract
E-learning, a modern method of education that utilizes electronic technologies such as
computers, mobile devices, and the internet, has experienced a significant surge in adoption
and usage in recent years. While it has the potential to reach every corner of the world, it also
creates an opportunity for time and resource wastage. In almost all cases students use the same
device for studying and for entertainment purposes. Being one click away from ever-addicting
social media, it is very difficult for students to stay focused on studying using digital devices
and not waste time on it. The issue is quite significant as online education will be practised
more and more in the future. In spite of that, detecting the on-screen activity of students is an
underexplored region of research, and to our best knowledge, no research takes protecting their
privacy into consideration. Therefore in this research, a privacy-preserving architecture is
proposed to detect whether students are utilizing their time on their computer or wasting it
while the user’s privacy is protected with federated learning. A dataset containing over 4000
screenshots of different activities of students is used to classify them into categories using
several pre-trained models where our proposed FedInceptionV3 achieves a state-of-the-art test
accuracy of 99.75%.
Existing System
As the fourth industrial revolution is taking place, the internet is becoming faster and
more accessible to everyone [1]. From education to professional work, digital devices have
become essential tools. In recent years, online education has experienced exponential growth,
breaking down geographical barriers and offering a flexible learning environment [2]. A study
conducted in 2017 revealed a consistent increase in the number of US students enrolling in at
least one online course for 14 consecutive years, with over 6 million students taking online
courses in 2016. This surge in popularity reflects a growing demand for accessible and flexible
education [3]. As e-learning continues to gain traction, its popularity is expected to further soar
in the upcoming years. The growth of online learning is predicted to excel in upcoming years,

2. and it is expected that the global e-learning market is projected to reach $325 billion by 2025
[4]. However, ensuring students’ engagement in online learning poses unique challenges. One
of the challenges associated with online learning is the potential for unwanted time and
resource wastage. Because students often use the same device for both studying and
entertainment purposes, this makes it difficult to maintain focus on academic tasks, as social
media and other distractions are only a click away. Even unintentionally, students tend to open
social media while studying using a digital device. As a result, they end up wasting valuable
time instead of dedicating themselves to their studies. S. Khan et al. found that social media
use had a negative impact on academic performance, particularly for students who spent more
time on social media [5]. This finding is consistent with research by Geot [6]. Another study,
conducted by u. et al. in the field of education, uses a cross-sectional survey design to gather
data from a sample of 379 students from four universities in the Khyber Pakhtunkhwa region
[7]. The results of the study indicate that the majority of the students perceived social media to
have a negative impact on their academic performance. The study also found that the amount
of time spent on social media was positively correlated with negative perceptions of its impact
on academic performance.
Drawback in Existing System
 Communication Overhead:
Federated learning involves frequent communication between the central server and
individual devices. This communication can result in increased latency, especially in
situations where network conditions are not optimal. This may affect the real-time
nature of certain e-learning activities.
 Heterogeneous Data Distribution:
Users in an e-learning environment may have diverse learning patterns, preferences,
and devices. If the data distribution across devices is highly heterogeneous, federated
learning may face challenges in converging to a robust and accurate global model.
 Resource Intensive Training:
Federated learning requires computation resources on both the client devices and the
central server. In resource-constrained environments, this may pose challenges,
affecting the overall efficiency and responsiveness of the e-learning system.

3.  User Resistance and Trust Issues:
Users may be resistant to sharing even aggregated and anonymized information
about their on-screen activities due to privacy concerns. Building and maintaining
trust in the federated learning system is essential for its successful adoption in e-
learning environments.
Proposed System
 The proposed system with FedInceptionV3 predicts 798 samples correctly out of 800
samples in the test sample and achieves a remarkable test accuracy of 99.75%.
 Proposed a more effective approach that involves analyzing the actual content displayed
on the screen, rather than relying on URL-based site tracking.
 One of the key strengths of our proposed system lies in the fact that the detection
process occurs entirely on the local device itself, eliminating the need for sample
uploads to a remote server.
 Proposed method with three state-of-the-art approaches in terms of accuracy, precision,
recall, and F1-score. The evaluation was conducted considering privacy concerns.
Algorithm
 Federated Learning:
Federated learning itself is a key algorithm in this context. It allows machine
learning models to be trained across decentralized devices. The federated learning
process typically involves three main steps: model initialization, local training on
individual devices using local data, and model aggregation to create a global model
without exchanging raw data.
 Homomorphic Encryption:
Homomorphic encryption allows computations to be performed on encrypted data
without decrypting it. In the context of federated learning, this can be used to perform
computations on locally held data without revealing the raw data to the central server.
This adds an extra layer of privacy protection.
 Secure Aggregation:
Secure aggregation protocols, such as federated averaging with secure aggregation,
ensure that the model updates from individual devices are combined in a way that
prevents the central server from learning specific details about any single user's data.

4. Advantages
 Data Security:
Federated learning incorporates privacy-enhancing technologies such as differential
privacy, homomorphic encryption, and secure multi-party computation. These
techniques ensure that user data remains secure and confidential during the model
training process, reducing the risk of data breaches.
 Improved Generalization:
By learning from a diverse set of user behaviors across different devices, federated
learning can lead to more generalized models. This is particularly beneficial in e-
learning, where users may have unique learning styles and preferences.
 Adaptability to Dynamic Environments:
E-learning environments can be dynamic, with users joining and leaving at different
times. Federated learning adapts well to these changes, allowing models to be
continuously improved without requiring centralized retraining every time there is a
change in the user population.
 Energy Efficiency:
Federated learning can be more energy-efficient compared to centralized
approaches, as the training process is distributed across devices. This is important for
resource-constrained devices, such as those used by students in e-learning scenarios.
Software Specification
 Processor : I3 core processor
 Ram : 4 GB
 Hard disk : 500 GB
Software Specification
 Operating System : Windows 10 /11
 Frond End : Python
 Back End : Mysql Server
 IDE Tools : Pycharm