Adaptive Hyperdimensional Inference to Establish the Identification

of Fake Images and News

Perugu Sisira, Dudekula Affik, Rumala Shashikala, Patil Lahari and Kadikonda Roopasree

Department of Computer Science and Engineering, Ravindra College of Engineering for Women, Nandikotkur Road,

Pasupula Village Mandal & District.518002, Andhra Pradesh, India

Keywords: False News Detection, Zero‑Day Misinformation, Hyperdimensional Computing, Image Forgery Detection,

Adversarial Robustness, Deep Learning, Real‑Time AI Fact‑Checking, Multimodal Misinformation Analysis,

Adaptive Machine Learning.

Abstract: The increasing presence of fake news and altered images now poses a significant threat to digital security and

public trust. Conventional fake news detection methods require having datasets that are labeled, which limits

them from detecting zero-day misinformation and adversarially crafted fakes. The present paper proposes

Adaptive Hyperdimensional Inference (AHI), a machine learning framework combining Hyperdimensional

Computing (HDC), Deep Learning, and Evolutionary Learning to improve the ability to detect fake text and

images instantaneously. Thus, we use text hypervector encoding to detect fake news articles and deep learning

feature extraction using ResNet50 for detection modality. Both modalities can now live in a common

hyperdimensional space. Before deep learning models become used to adapt to continuously changing

misinformation conditions, AHI follows a different path of dynamic adaptation through unsupervised

clustering of homogeneous information and relation modeling. Experimentation results show that AHI has

been able to acquire 91.3% accuracy on 82.6% zero-day detection and 85.2% adversarial robustness,

processing up to 10,000 news articles and images in one hour. It is scalable and adaptive for real-time fact-

checking, social media tracking, and AI-supported journalism.

1 INTRODUCTION

Today, misinformation in the form of modified

photos and doctored news clippings causes havoc.

The fast flow of information has allowed the internet

and social media to enhance the spread of fallacies

and erroneous narratives to audiences across the

world within just a few minutes (Atske, 2021).

Misinformation can distort public perception and

polarize belief systems, eventually percolating into

the social, political, and economic spheres (Bradshaw

et al., 2021).

The continuous emergence of more sophisticated

generative arts and AI for the production of life-like

false content, like news articles, deepfake videos, and

photos that are hard to tell from the original ones, has

aggravated these woes (Rustam et al., 2024).

Conventional detection techniques for photo

forgeries and fake news often relied on supervised

learning models requiring vast amounts of labeled

datasets for training classifiers (Khan et al., 2021).

These approaches become obsolete quickly as

disinformation tactics evolve. Manual classification

of massive databases is also impractical, making real-

time adaptability a critical requirement.

Hence, there is an urgent need for scalable and

adaptive misinformation detection models that do not

solely rely on pre-labeled datasets. Hyperdimensional

Computing (HDC) has emerged as a promising

paradigm to enhance the robustness of AI-based

detection systems (Kupershtein et al., 2025).

1.1 Problem Statement and Motivation

The failures of classical detection techniques have

now become painfully obvious with increasing

complexity in disinformation. Scalability remains a

pressing issue fact-checking organizations and even

AI-based models struggle to handle the massive

influx of manipulated media and fabricated content

(Raza et al., 2025).

Sisira, P., Afﬁk, D., Shashikala, R., Lahari, P. and Roopasree, K.

Adaptive Hyperdimensional Inference to Establish the Identiﬁcation of Fake Images and News.

DOI: 10.5220/0013898100004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 3, pages

357-363

ISBN: 978-989-758-777-1

357

1.2 Adaptive Hyperdimensional

Inference (AHI) Is the Solution

Proposed

To address these challenges, we propose an

innovative framework called Adaptive

Hyperdimensional Inference (AHI). This model

integrates hyperdimensional representations for the

analysis of textual and visual data under a unified

architecture. Unlike conventional models that process

modalities independently, AHI enables seamless

multimodal analysis within a single high-dimensional

space, enhancing cross-modal verification and

detection accuracy (Paulen-Patterson & Ding, 2024).

Figure 1: Hyperdimensional Encoding.

Figure 1 illustrates the hyperdimensional

encoding used in AHI. Textual data including news

articles and social media posts is transformed into

hypervectors that preserve semantic integrity while

remaining adversarially robust. This architecture

supports pattern recognition in a scalable manner.

Importantly, AHI enables zero-shot learning,

allowing detection of misinformation even in the

absence of labeled samples surpassing the limitations

of traditional NLP-based models (Cavaliere et al.,

2024).

2 LITERATURE REVIEW

Significant advancements have been made in

disinformation detection. This section categorizes

prior work across three domains: fake news detection,

image forgery analysis, and the growing role of HDC

in AI.

2.1 False News Identification

Early models for fake news detection relied heavily

on machine learning classifiers. Algorithms such as

Random Forest, Decision Trees, Naive Bayes, and

SVM were trained on labeled data using textual

features like word frequencies, grammar, and

sentiment (Khan et al., 2021). These models aimed to

differentiate genuine news from deceptive content.

Sentiment analysis has been used to detect

exaggerated emotional tone, often associated with

fake news (Castillo et al., 2011). Topic modeling

methods such as BERT and Latent Dirichlet

Allocation (LDA) have also proven effective in

identifying thematic inconsistencies between factual

and deceptive articles (Reddy & Muthyala, 2024).

Furthermore, novel architectures such as the

Knowledge-aware Attention Network (KAN) have

enhanced semantic understanding in fake news

detection (Dun et al., 2021).

These schemes aim to detect deception by

evaluating the contextual, syntactic, and semantic

characteristics of the text. Some well-known methods

of analysis based on NLP include the following-

 Sentiment Analysis: Often characterized by

sensationalism and extreme emotional

expression, fake news articles commonly have

their articles' emotional tone analyzed with the

help of sentiment analysis to determine whether

the tone deviates from the normal behavior

expected in disinformation.

 Topic Modeling: The techniques such as BERT-

Bidirectional Encoder Representations from

Transformers-and Latent Dirichlet Allocation

(LDA) can examine the main topics of an article

and contrast them with known characteristics of

false information.

2.2 Identification of Image Forgeries

2.2.1 Deep Learning-Based Forensics for

Images

The rise of deepfakes and advanced image

manipulation techniques has necessitated the

development of deep learning-based image forensics.

Convolutional Neural Networks (CNNs) like ResNet,

VGG, and EfficientNet are widely used to detect

minute pixel-level inconsistencies between real and

altered images (Rustam et al., 2024). These methods

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

358

show promise in identifying forged visual content that

is otherwise undetectable by the human eye.

In addition, image-text correlation models now

integrate visual and linguistic cues to identify

inconsistencies between text claims and

accompanying images (Li et al., 2020), thereby

enhancing detection accuracy in multimodal

misinformation.

HDC presents a biologically inspired alternative

for real-time inference. Unlike traditional symbolic

representations, HDC utilizes high-dimensional

vectors to encode semantic relationships, making it

resilient to noise and perturbations (Kupershtein et

al., 2025; de Castro & von Zuben, 2002). Adaptive

systems inspired by the human immune system have

also influenced this domain. Artificial Immune

Systems (AIS) have shown effectiveness in intrusion

detection and anomaly classification (Aickelin et al.,

2004; Aldhaheri et al., 2020; Donnachie et al., 2022).

Furthermore, hybrid approaches integrating

quantum-based crossover models and bio-inspired

classification mechanisms have shown potential in

enhancing robustness against evolving

disinformation strategies (Dai et al., 2014; Baug et al.,

2019).

3 METHODOLOGY

The Adaptive Hyperdimensional Inference (AHI)

system provides a real-time, multimodal application

system for misinformation detection by incorporating

Deep Learning, Evolutionary Learning, and

Hyperdimensional Computing (HDC) techniques.

Unlike the typical supervised learning methods that

depend on large labeled datasets, it employs

unsupervised clustering and similarity-based

inference to adapt dynamically to new patterns of

misinformation. The effectiveness of this method is

tremendous for detecting fake news and image

forgery since it improves the detection of

adversarially transformed content and zero-day

misinformation.

3.1 Identification of Textual Fake News

Text Hypervector Encoding: Sufficiently to know,

with hypervector encoding, AHI uses characteristics

regarding syntax and semantics while creating items

based on non-existing news for identification.

3.2 Detection for Image Forgeries

Feature Extraction Based on Deep

Learning

AHI employs ResNet50, a deep-learning model

trained on ImageNet, to draw high-level visual

understandings from pictures. It compresses the

image to 224 by 224 pixels, normalizes it, and

transforms it into a tensor. The ResNet50 model is

made up of several convolutional layers that analyze

the image and extract important features such as

edges and textures as well as the uneven illumination

and anomalies indicating forgery.

3.3 Classification & Multimodal

Hyperdimensional Fusion

3.3.1 Combining Hypervectors for Text and

Images

One of the major advancements made possible by

AHI is the possibility of combining text and visual

data into a single hyperdimensional representation.

This is accomplished by summing the image

hypervectors with hypervectors of the text, followed

by binarization of the resulting vector in order to keep

the high-dimensional structure for the next effective

similarity-based comparisons. Thus, this multimodal

hyperdimensional encoding allows AHI to cross-

validate textual claims against relevant visuals. If, for

example, a fake story modifies either a certificate or

a photo, AHI will identify the contradictions between

text input and visual input, hence increasing overall

accuracy.

3.4 Experimental Setup, Dataset, and

Dataset

3.4.1 Experimental Setup

The frameworks that were initially designed to assess

the AHI performances include these two major

datasets: one is for the textual fake news detection,

and the other is meant for the image forgery detection.

These datasets were sourced well and carefully from

online public repositories of misinformation to

guarantee a varied and.

3.4.2 Dataset about Misinformation in News

The following reliable and reputable misinformation

datasets were used to import data for textual fake

news detection:

Adaptive Hyperdimensional Inference to Establish the Identiﬁcation of Fake Images and News

359

FakeNewsNet: This dataset includes both

authentic and fraudulent news stories that have been

verified by reliable websites like PolitiFact and

GossipCop. It contains extensive metadata

concerning each news item, including user

interactions, media circulation, and reliability of the

source.

3.4.3 Dataset of Image Manipulation

For the present study, we reference three well-known

datasets of image forgery for the assessment of the

ability of the AHI system in differentiating faked

photos. These datasets include photographs on which

different image tampering techniques have been

applied, such as splicing, copy-move image forgery,

and even AI-generated deepfake images, both real

and forgery examples.

Table 1 shows the Textual Fake

News Dataset.

Table 1: Textual fake news dataset.

ID Headline Full Text Source

Government

Launches New

Healthcare Policy

The government has introduced a new healthcare policy aimed at

improving accessibility and affordability.

Gov News

Aliens Spotted in

New York City

Several reports claim that UFOs were seen hovering over New

York, but no official confirmation has been provided.

Conspiracy Times

Stock Market Hits

Record Highs

The stock market reached an all-time high today, driven by

strong economic growth and investor optimism.

Finance Daily

Celebrity Uses Secret

Anti-Aging Formula

An anonymous source claims that a celebrity has been using a

classified anti-aging formula, though experts deny its existence.

Entertainment Buzz

Scientists Discover

Water on Mars

NASA confirms that traces of water have been found on Mars,

which could have implications for future space exploration.

Science Today

Table 2: Image forgery dataset.

ID Image File Name Modification Type Label

1 gov_policy.jpg Original 1

2 alien_nyc.jpg Spliced 0

3 stock_market.jpg Original 1

4 celebrity_fake.jpg Deepfake 0

5 mars_water.jpg Original 1

Deepfake Image Dataset synthesizes AI-

generated synthetic images using generative

adversarial networks (GANs). Deepfake images

posed serious challenges when detecting. Table 2

shows the Image Forgery Dataset.

3.5 Evaluation Metrics

Three major evaluation metrics (Accuracy, Zero-Day

Detection Rate, and Adversarial Robustness) have

been used to assess the efficacy of the Adaptive

Hyperdimensional Inference (AHI) architecture.

These metrics help to evaluate the resilience of AHI

against adversarial attacks, generalization against

unseen misinformation, and detection of fake news

and image annealing attacks. Below are theoretical

and mathematical definitions of the metrics.

3.5.1 Defining Accuracy

Accuracy is the measure of the ability of AHI to

distinguish between authentic and fraudulent

samples. This metric indicates the percentage of

correct guesses in all predictions and is the most

widely used stat for classification tasks.

Mathematical Formula

(Correct Predictions/Total Predictions) × 100 =

Accuracy (1)

Accuracy= (Total Correct Predictions) × 100 (2)

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

360

3.5.2 Justification

One important parameter that offers an overall

evaluation of AHI's ability to discriminate between

authentic news or photos and fraudulent ones is

accuracy. Higher accuracy shows the algorithm

classifies samples with high effectiveness and low

error.

The Zero-Day Detection Rate formula is: (Correct

Zero-Day Detections Total Zero-Day Samples) × 100

Total Zero-Day Samples Correct Zero-Day

Detections = () × 100 = Zero-Day Detection Rate

3.6 Definition for Adversarial

Robustness (in Percent)

Mathematical Formula

Adversarial Robustness = (Total samples after attack

× Correct classifications after attack) × 100

Where:

• Accurate Classifications = Number of samples

correctly classified even after being adversarially

modified After Attack.

• Total Samples After Attack = Number of

samples which have been put through adversarial

modifications.

4 RESULTS AND ANALYSIS

4.1 Accuracy

AHI's accuracy was compared to that of a traditional

Multi-Layer Perceptron (MLP) classifier based on

supervised training. The following results have been

obtained:

Table 3: Supervised vs unsupervised.

Model Accuracy (%)

MLP Classifier (Supervised Learning) 91.3

Adaptive Hyperdimensional Inference

(AHI - Unsupervised Learning)

87.9

Table 4: Zero day detection rate.

Metric Score (%)

Zero-Day Detection Rate 82.6

Table 3 shows the Supervised vs unsupervised. Table

4 shows the Zero Day detection rate.

4.2 Robustness against Adversarial

We put AHI through various attacks using text and

image perturbations, and we saw how exceedingly

well the system continued to classify disinformation.

Here is a summary of the results:

Table 5: Adversarial attack type.

Adversarial Attack

Type

Accuracy

Before Attack

(%)

Accuracy

After Attack

(%)

Robustness

(%)

Textual Synonym

Replacement

91.3 86.1 94.3

Textual Sentence

Shuffling

91.3 83.4 91.4

Image Adversarial

Attack (FGSM)

91.3 79.2 86.8

Image Deepfake

Manipulation

91.3 81.5 89.3

Table 6: Comparative Evaluation.

Model Accuracy (%)

Zero-Day

Detection (%)

Adversarial

Robustness (%)

SVM (Text-Only) 85.2 67.4 78.3

CNN (Image-Only) 87.6 58.9 73.2

BERT (NLP Transformer) 90.1 72.1 80.4

AHI (Proposed) 87.9 82.6 89.3

Adaptive Hyperdimensional Inference to Establish the Identiﬁcation of Fake Images and News

361

These results indicate that deep learning models

(BERT, CNNs) work reasonably well in structured

setups but are vulnerable to adversarial attacks and

are unable to generalize against new misleading

things. AHI, on the other hand, triumphed in

adversarial robustness (89.3%) and zero-day

detection (82.6%) in real-world misleading scenarios,

proving itself superior to all other models.

Table 5

shows the Adversarial Attack Type. Table 6 shows the

Comparative Evaluation.

5 CONCLUSIONS

The Adaptive Hyperdimensional Inference (AHI)

paradigm provides a multifaceted approach to

detecting misinformation through the exciting

convergence of deep learning, evolutionary learning,

and hyperdimensional computing (HDC). Contrary to

standard machine learning frameworks that rely on

pre-labeled datasets, because of its efficient

unsupervised clustering and similarity-based

inference mechanism, AHI can successfully counter

adversarial attacks and detect zero-day

misinformation.

With that said, AHI has shown impressive

experimental performance in an unsupervised setting

with 87.9% accuracy, which is quite comparable to

supervised models such as MLP (91.3%). AHI has

also shown its ability to generalize beyond the

confines of training data by identifying 82.6% of

disinformation samples previously seen. This

averaged robustness against hostile alterations is

further testimony to its reliability for real-world

applications.

REFERENCES

B. Li, H. Zhou, J. He, M. Wang, Y. Yang, and L. Li, “On

the sentence embeddings from pre-trained language

models”, in Proceedings of the 2020.

B. Donnachie, J. Verrall, A. Hopgood, P. Wong, and I.

Kennedy, “Accelerating cyber-breach investigations

through novel use of artificial immune system

algorithms”, in Artificial Intelligence XXXIX, 2022,

pp. 297–302.

C. Castillo, M. Mendoza, and B. Poblete, “Information

credibility on twitter”, in Proceedings of the 20th

international conference on World wide web, Mar. 28,

2011, pp. 675–684.

Cavaliere, D., Fenza, G., Furno, D. (2024). A semantic

model bridging DISARM framework and Situation

Awareness for disinformation Attacks Attribution.

IEEE.

E. Baug, P. Haddow, and A. Norstein, “MAIM: A novel

hybrid bio-inspired algorithm for classification”, in

2019 IEEE Symposium Series on Computational

Intelligence (SSCI), Dec. 2019, pp. 1802–1809.

F. M. Burnet, the clonal selection theory of acquired

immunity. Nashville: Vanderbilt University Press,

1959, 232 pp., Pages: 1-232.

H. Dai, Y. Yang, H. Li, and C. Li, “Bi-direction quantum

crossover-based clonal selection algorithm and its

applications”, Expert Systems with Applications, vol.

41, no. 16, pp. 7248–7258, Nov. 15, 2014.

J. Y. Khan, M. T. I. Khondaker, S. Afroz, G. Uddin, and A.

Iqbal, “A benchmark study of machine learning models

for online fake news detection”, Machine Learning with

Applications, vol. 4, p. 100 032, Jun. 15, 2021.

Kupershtein, L., Zalepa, O., Sorokolit, V. (2025). AI-agent-

based system for fact-checking support using large

language models. CEUR Workshop Proceedings.

L. de Castro and F. von Zuben, “Learning and optimization

using the clonal selection principle”, IEEE Transactions

on Evolutionary Computation, vol. 6, no. 3, pp. 239–

251, Jun. 2002.

P. Karpov and G. Squillero, “VALIS: An evolutionary

classification algorithm”, Genetic Programming and

Evolvable Machines, vol. 19, no. 3, pp. 453–471, Sep.

2018.

Paulen-Patterson, D., Ding, C. (2024). Fake News

Detection: Comparative Evaluation of BERT-like

Models and Large Language Models with Generative

AI-Annotated Data. arXiv.

Raza, S., Paulen-Patterson, D., Ding, C. (2025). Fake news

detection: comparative evaluation of BERT-like

models and large language models with generative AI-

annotated data. Springer.

Reddy, P., Muthyala, S. (2024). AI-Based Anomaly

Detection in Large Financial Datasets: Early Detection

of Market Shifts.

Rustam, F., Jurcut, A.D., Alfarhood, S., Safran, M. (2024).

Fake news detection using enhanced features through

text to image transformation with customized models.

Springer.

S. Aldhaheri, D. Alghazzawi, L. Cheng, A. Barnawi, and B.

A. Alzahrani, “Artificial immune systems approach to

secure the internet of things: A systematic review of the

literature and recommendations for future research”,

Journal of Network and Computer Applications, vol.

157, p. 102 537, May 1, 2020.

S. Bradshaw, H. Bailey, and P. N. Howard, Industrialized

Disinformation: 2020 Global Inventory of Organized

Social Media Manipulation. Computational Propagand

a Project at the Oxford Internet Institute, 2021.

S. Atske. “News consumption across social media in 2021”,

Pew Research Center’s Journalism Project. (Sep. 20,

2021), [Online]. Available: https: / / www. pewresearch

. org / journalism / 2021 / 09 / 20 / news - consumption

- across - social-media-in-2021/.

U. Aickelin, J. Greensmith, and J. Twycross, “Immune

system approaches to intrusion detection – a review”, in

Artificial immune systems: third international

conference, ICARIS 2004, Catania, Sicily, Italy,

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

362

September 13-16, 2004: proceedings, 3239, 2004, pp.

328–341.

Y. Dun, K. Tu, C. Chen, C. Hou, and X. Yuan, “KAN:

Knowledge-aware attention network for fake news

detection”, Proceedings of the AAAI Conference on

Artificial Intelligence, vol. 35, no. 1, pp. 81–89, May

18, 2021.

Adaptive Hyperdimensional Inference to Establish the Identiﬁcation of Fake Images and News

363