Internet and Telecommunication Fraud Prevention Analysis based on Deep Learning

Figures & data

Figure 1. New internet and telecommunication fraud risk factors.

Figure 2. Model structure.

Table 1. Data distribution and labeling.

Dataset	Description	Data labeling status	Number of data items
dataset1	Police notification data obtained from Internet channels-Part1	The dataset is labeled with 12 types of tags such as fraudulent tools and modus operandi	436
dataset2	Police notification data obtained from Internet channels-Part2	The dataset is labeled with 11 types of tags such as fraudulent tools, victim geography, fraudulent property	1294
dataset3	Fraud case from media data	The dataset is labeled with 19 types of tags such as modus operandi, targeted audience, and payment method	1166
dataset4	Case document data from multiple judgement website	The dataset is labeled with 23 types of tags such as fraudulent tools, modus operandi, targeted audience, victim geography, and fraudulent property	608
Total			3504

Table 2. The labeled data distribution of risk factor.

Risk Factor Category	Risk factor level 2 category	Data volume
Risk factors of fraudulent tools	Phone	367
	Network	1241
Risk factors of modus operandi	Identity impersonation scam techniques	81
	Lure-type fraudulent practices	249
	Game type fraudulent techniques	169
	Shopping scams	85
	Fictitious risk scamming techniques	22
	Phishing, Trojan horse-type fraudulent practices	83
	Loan financial fraudulent practices	349
	Daily life consumption fraudulent techniques	180
	Subsidized tax refund scam techniques	27
	Other new types of fraudulent practices	323
Risk factors of targeted audience	Age	74
	Gender	188
	Career	48
	Disease Conditions	12
Risk factors of victim geography	East China	140
	South China	127
	North China	90
	Central China	132
	Northeast China	81
	Northwest China	72
	Southwest China	104
Risk factors of fraudulent property	Defrauded property	202
	Amount defrauded	665
Risk factors of payment method	Bank Card Transfer	169
	Red Envelop	47
	Scan the code	29
	QQ transfer	43
	WeChat transfer	87

Table 3. Parameter settings.

Parameter	Value	Parameter	Value
maximum sequence length	256	learning_rate	0.00005
batch_size	16	epoch	20
drop_out	0.1	warmup_proportion	0.1
num_tuning_layers	12	num_BiLSTM_layers	2
Optimizer	AdamW	Activation function	GELU

Table 4. Baseline experiment.

Model	Task	Data source
Model 1	Training in extraction of factors of fraudulent tools	dataset1, dataset2, dataset4
Model 2	Training in extraction of factors of modus operandi	dataset1, dataset3, dataset4
Model 3	Training in the extraction of factors of targeted audience	dataset3, dataset4
Model 4	Training in extraction of factors of victim geography, fraudulent property	dataset2, dataset4
Model 5	Training in extraction of factors of payment method	dataset3

Table 5. Fusion model experiment.

Model	Task	Data source
Fusion Model	Training in extraction of fraudulent tools, modus operandi, targeted audience, victim geography, fraudulent property, and payment method	dataset1, dataset2, dataset3, dataset4

Table 6. Resource usage during training.

trainingResource_GPU	Experiment	trainingTime_GPU	training_duration
One GPU, 11 GB of video memory	baseline experiments (Serial execution of 5 training tasks)	92 min	92 min
	Fusion experiments (Only 1 training mission required)	29 min	29 min
	Time optimization ratio	68.5%	68.5%
5 GPUs, 55 GB of video memory	baseline experiments (Parallel 5 training tasks）	94 min	26 min
	Fusion experiments (Can be accelerated with multi-card parallel training)	37 min	8 min
	Time optimization ratio	60.6%	69.2%

Table 7. Resource usage during model inference.

Experiment	inferenceResource_GPU_idle	inferenceResource_GPU_busy	DPS
baseline experiments (5 models)	7875 MB	8991 MB	48
Fusion experiments (1 model)	1575 MB	2691 MB	235
Resource optimization ratio	80%	70.10%	-

Table 8. Effect comparison.

Experiment	precision	recall	f1	Increase
random sample (formula 1)	80.44%	76.95%	78.66%	-
smooth sample (formula 2–3)	81.25%	79.20%	80.21%	1.56%

Figure 3. Tensorboard graph.

Table 9. Effect comparison.

	precision	recall	f1	Increase
baseline loss formula without adjustment (formula 4)	81.62%	77.68%	79.60%	-
loss formula with smooth loss weight (formula 5–7)	82.71%	77.71%	80.13%	0.53%
loss formula with loss weight of fusion factor (formula 8–11)	83.73%	78.88%	81.23%	1.63%

Figure 4. Tensorboard graph.