An Improved Image Spam Classification Model Based on Deep Learning Techniques

Table 4

Accuracy comparison on Dredze et al. [10] and ISH datasets [11].

	Dataset 1 [10]	Dataset 2 [11]
Our proposed CNN model	99.44	99.77
Sriram et al. (DCNN) [30]	97.30	99.80
Annadatha and Stamp (SVM) [4]	—	97.00
Chavda et al. (SVM) [5]	98.00	97.00
Qian et al. (SVM with Gaussian kernel) [26]	97.90	98.30
Gao et al. (ISH) [11]	—	94.94
Yuan and Zhang (multifeatures fusion method) [28]	95.00	—
Das et al. (multiple classifier) [9]	98.00	—
Kumar and Biswas (image texture features) [25]	98.50	98.56
Dredze and Gevaryahu [10]	98.00	—
Shen et al. (comprehensive visual modeling) [47]	96.80	—
Wang et al. (low-level image feature) [19]	97.00	—
Al-Duwairi et al. (Base64 encoding) [23]	99.00	—
Al-Duwairi et al. (texture analysis) [24]	98.55	98.10
Liu et al. (multilayer spam filter) [22]	94.30	94.30
Gupta et al. (low-level and metadata features) [20]	93.30	—
Xu et al. (K-labels propagation model) [27]	90.00	—
Soranamageswari and Meena (ANN) [16]	92.82	—
Kumaresan et al. (SVM and PSO) [21]	90.00	—
Yang et al. (multimodal fusion) [29]	92.64	—