Figure 1.
Figure 2.
Figure 3.
Figure 4.
Figure 5.
Politifact Experimental Results
| Model | Precision | Recall | F1-Score | Accuracy |
|---|---|---|---|---|
| GB | 75.37 | 75.67 | 75.32 | 74.56 |
| XGBOOST | 75.24 | 75.25 | 75.56 | 76.14 |
| LR | 81.24 | 81.35 | 81.40 | 81.46 |
| DT | 77.65 | 77.83 | 77.75 | 76.90 |
| SVC | 83.26 | 83.41 | 83.36 | 83.45 |
| LSTM | 84.24 | 84.65 | 84.43 | 84.32 |
| Bi-LSTM | 82.92 | 83.87 | 83.57 | 83.14 |
| CNN Bi-LSTM | 82.10 | 81.36 | 81.78 | 81.67 |
| BERT | 87.05 | 87.19 | 88.07 | 88.81 |
| DistilBERT | 86.08 | 86.29 | 86.07 | 86.45 |
| DeBERTa | 84.19 | 84.56 | 84.07 | 84.68 |
| GPT-3 Ada | 91.30 | 96.18 | 93.68 | 91.98 |
| GPT-3 Baggage | 91.75 | 96.01 | 92.58 | 91.54 |
| GPT-3 Curie | 91.85 | 94.66 | 93.23 | 91.50 |
| GPT-3 Davinci | 92.65 | 94.56 | 92.87 | 90.87 |
Inference Latency and Deployment Suitability of GPT-3 Variants
| Model | Latency (ms) | Cost/1K Tokens | Accuracy (Avg) | Edge Suitability |
|---|---|---|---|---|
| Ada | 50–100 | $0.0004 | 93.84% | High |
| Babbage | 100–200 | $0.0005 | 92.88% | Moderate–High |
| Curie | 150–300 | $0.002 | 93.33% | Moderate |
| Davinci | 300–500 | $0.02 | 93.22% | Low |
Comparison with State of the Art literature
| Authors | Contribution | Methodology |
|---|---|---|
| Cui et al. [6] | Proposed a framework to integrate various features for fake news and significant improvement in the performance | Multi-modal embedding with sentiment awareness in fake news detection |
| Khan et al. [9] | Proposed a novel approach for explainable fake news detection | Using pyramidal co attention networks |
| Shadi et al. [14] | Improving the performance of the fake news detection | Enhancing BERT through attention mechanisms |
| Chen et al. [15] | Exploring detection of fake news on Covid-19 | Explored various deep learning algorithms |
| Wazid et al. [16] | Proposed an intrusion detection machine learning model for healthcare applications in industry 5.0 | Involves ensembling approaches through various machine learning classifiers |
| Li et al. [17] | Proposed STAR-RIS-CR-NOMA in Consumer IoT Networks for Industry 5.0 Resilience | Presented a novel approach named STAR-RIS-NOMA |
| Hussein et al. [18] | Introducing a novel CSI compression model that approximates sufficient statistics functions for individual channel matrices for industry 5.0 | High degree of synchronization through the base station (BS) to obtain instant channel state information (CSI) |
| He et al. [4] | Proposed a reinforcement learning approach for the detection of the fake news in social media | A reinforcement framework via counter responses generation |
| Ours | AI based zero trust security for GenAI systems in industry 5.0 that is human centric and sustainable | Employing GPT-3 for misinformation detection |
Gossipcop Experimental Results
| Model | Precision | Recall | F1-Score | Accuracy |
|---|---|---|---|---|
| GB | 82.35 | 82.68 | 78.53 | 81.75 |
| XGBOOST | 80.24 | 82.43 | 80.46 | 81.68 |
| LR | 84.57 | 84.46 | 83.24 | 84.40 |
| DT | 77.45 | 77.28 | 77.83 | 77.48 |
| SVC | 84.35 | 85.63 | 84.67 | 84.98 |
| LSTM | 80.25 | 82.65 | 80.38 | 79.78 |
| Bi-LSTM | 79.32 | 79.54 | 79.21 | 78.56 |
| CNN Bi-LSTM | 81.34 | 81.54 | 81.67 | 81.32 |
| BERT | 77.10 | 89.65 | 82.43 | 85.68 |
| DistilBERT | 78.45 | 89.09 | 82.34 | 85.70 |
| DeBERTa | 78.24 | 89.43 | 82.21 | 85.45 |
| GPT-3 Ada | 96.48 | 97.86 | 97.18 | 95.69 |
| GPT-3 Baggage | 95.68 | 97.13 | 96.19 | 94.21 |
| GPT-3 Curie | 95.89 | 97.56 | 96.95 | 95.15 |
| GPT-3 Davinci | 96.34 | 97.66 | 97.01 | 95.56 |
Performance comparison with Literature Survey
| Politifact | Gossipcop | |
|---|---|---|
| Dong et al. [5] | 82.39% | 94.86% |
| Cui et al. [6] | 77.24% | 80.42% |
| Furukawa et al. [7] | 85.1% | 90% |
| Salama et al. [8] | 83.93% | 83.82% |
| Shu et al. [35] | 90.4% | 80.8% |
| Proposed | 91.98% | 95.69% |