An Enhanced Privacy-Preserving Federated Learning Framework Using Attribute-Based Encryption for Smart IoT Systems

Table 1. Quick summary of the different existing works

Lin et al. [27] introduced SACM, an attribute-relied Secure Access Control Mechanism for IoT-Health leveraging FL. The framework innovatively correlated users’ social attributes with trust levels based on social influences using graph convolutional networks. Access permissions were determined through trust thresholds specific to each occupation, with federated deep learning optimizing control parameters. Experimental results showed effective access control. However, the system’s performance in highly dynamic social networks required additional investigation.

Islam and Madria [28] proposed a revocable collusion-resistant ABE scheme supporting unlimited user revocation without affecting non-revoked users’ secret membership keys. The framework extended to DMG-SDS, enabling dynamic multi-group operations while maintaining key integrity. Performance assessment showed significant advantages over contemporary schemes in multi-group data sharing scenarios. The approach effectively addressed security requirements while maintaining practical implementation capabilities. Meanwhile, the system’s efficiency with extremely large user groups might benefit from further optimization. Table 1 highlights the summary of distinct existing procedures.

1.png

Figure 1. Proposed framework for the ABE-HCL enabled FL model

2.png

Figure 2. LSTM structure

The recommended hybrid learning model integrates LSTM with the Whale Optimization approach. LSTM consists of three key components: the input gate (I.G), the forget gate (F.G), the cell input (V.I), and the output gate (D.G). Typically, LSTM is a memory-driven neural network(NN) designed to retain values after each loop. Given that x_t represents the current input and h_t denotes the output at time step t, with h_t−1 as the previous output, the cell input state is V_t, and the cell output state is L_t (with its preceding state L_t-1, the gates' states are represented by k_t, T_f, and T₀. The structure of LSTM operates such that both L_t and h_t are transmitted to the subsequent NN layer in the Recurrent Neural Network (RNN). The LSTM mechanism merges the output of the prior unit with the recent input state, where the output and forget gates facilitate updates to the memory. To determine G_t and h_t, the pursuing equations are utilized.

$I. G: k_t=\theta\left(L_l^i \cdot D_t+L_h^i \cdot e_{t-1}+s_i\right)$                        (1)

$F. G: T_f=\theta\left(G_l^f . D_t+L_h^f . e_{t-1}+s_f\right)$                        (2)

$D \cdot G: T_o=\theta\left(L_l^0 \cdot D_t+L_h^o \cdot e_{t-1}+s_o\right)$                        (3)

$V . I: \widetilde{T_C}=\tanh \left(L_l^C . D_t+L_h^C . e_{t-1}+s_C\right)$                        (4)

The weight matrices among the input gates and output layers are denoted as $L_l^0, L_l^f, L_l^i, L_l^C$, while the weight criteria among the hidden and input layers are represented by $L_l^0, L_l^f, L_l^i, L_l^C$. The bias vectors are labelled as $s_i, s_f, s_o, s_C$ and the hyperbolic function tanh is applied. The output state of the cell is ascertained by the pursuing formulas:

$T_C=k_t * \widetilde{T_C}+T_f * T_{t-1}$                        (5)

$e_t=T_o * \tanh \left(T_C\right)$                        (6)

The equation above yields the final output score.

4.3 Extended LSTM model –Its working mechanism

To elevate the storage capacity of LSTMs, the memory cell is extended from a scalar $c \in R$ to a matrix $C \in R^{\text {dxd }}$, allowing for retrieval via matrix multiplication. At a given time, $t$, a pair of vectors-a key $k_t \in R^d$ and a value $v_t \in R^d$ stored, following the terminology used in Transformers. Subsequently, at time $t+\tau$, the value $v_t$ is acquired using a query vector $q_{t+\tau} \in R^d$. This mechanism aligns with the framework of Bidirectional Associative Memories (BAMs). The process employs the covariance update rule to encode the key-value pairs effectively.

$C_t=C_{t-1}+v_t k_t^T$                        (7)

We presume a layer normalization step is performed prior to projecting inputs into key and value spaces, ensuring these projections have a mean of zero. The rule for updating the covariance matrix is designed to optimize the separability of retrieved binary vectors. This optimal separability corresponds to achieving the highest possible signal-to-noise ratio. Enhanced separability can be achieved by restricting acquired to pairwise interactions and accepting the quadratic computational complexity associated with attention mechanisms.

Building on this foundation, we embed the covariance update rule within the LSTM architecture. In this setup, the forget gate $f_t$ acts as a decay factor, controlling how much of the previous memory is retained. The input gate $i_t$ regulates the learning rate by controlling the flow of new information into memory. The output gate $O_t$ adjusts the influence of the current memory state on the output. These gates are standard in LSTM architectures and are integrated here to manage the dynamics of memory update and retrieval.

Within this matrix memory framework, the normalizer state $n_t$ is defined as a weighted summation of key vectors, with weights determined by the input gate and the cumulative influence of all subsequent forget gates. This normalizer state effectively captures the dynamics of gate strengths. Given that the dot product among the query and the normalizer state can approach zero, taking the magnitude of this dot product and setting a minimum threshold (commonly 1.0) to ensure stability. Consequently, the forward propagation in the mLSTM model proceeds as:

$C_t=f_t C_{t-1}+i_t v_t k_t^T  \\  \quad  \text {Cell state}$                         (8)

$n_t=f_t n_{t-1}+i_t k_{-} t \\  \quad  \text {Normalizer state}$                          (9)

$\begin{gathered}h_t=o_t \odot \widetilde{h_t}, \widetilde{h_t}=C_t q_t / \max \left\{\left|n_t^T q_t\right|, 1\right\} \\ \text { Hidden state }\end{gathered}$                         (10)

where, the intermediate vectors are computed as:

$\begin{gathered}&q_t=W_q x_t+b_q\\&\text { Query input }\end{gathered}$                        (11)

$\begin{gathered}&k_t=\frac{1}{\sqrt{d}} W_k x_t+b_k\\&\text { Key input }\end{gathered}$                     (12)

$\begin{gathered}&v_t=W_v x_t+b_v\\&\text { Value input }\end{gathered}$                         (13)

Gate activations are calculated using:

$\begin{gathered}&i_t=\exp \left(\tilde{l_t}\right), \tilde{l_t}=w_i^T x_t+b_i\\&\text { Input gate }\end{gathered}$                        (14)

$\begin{gathered}&f_t=\sigma\left(\tilde{f}_t\right) O R \exp \left(\widetilde{f}_t\right), \widetilde{f}_t=w_f x_t+b_f\\&\text { Forget gate }\end{gathered}$                         (15)

$\begin{gathered}&o_t=\sigma\left(\tilde{o_t}\right), \tilde{o_t}=W_o x_t+b_o\\&\text { Output gate }\end{gathered}$                         (16)

The mLSTM framework, similar to the traditional LSTM, accommodates multiple memory cells. In the context of mLSTM, having multiple heads is synonymous with having multiple cells due to the absence of memory integration.

3.png

Figure 3. Entire working process for the Ashera CAT swarm optimization

(2) For every instance, randomly choose a set number of CDC dimensions to undergo mutation. Additionally, randomly increase or decrease the SRD values from the existing values, which will interchange the previous locations, as described in Eq. (17).

$x\left(new_{-}cat\right)=(1+rand + SRD)*x\left(old_{-}cat\right)$                          (17)

where, x(new_Ashera CAT) is new Ashera CAT’s new location, x(old_Ashera CAT) is Ashera CAT’s initial location and rand is random interval of time among [0,1].

The Fitness Function (FF) is calculated as per the following expression (18), and the candidate location is chosen which relies on the highest probability corresponding to the Fitness Function value. This approach selects the position with the highest FF, ensuring the optimal candidate is selected based on its fitness ranking.

                $P(i)=\mid\left(F F(i)-F F(b) \mid /\left(F F_{\max }-F F_{\min }\right)\right.$          (18)

The fitness of the recent Ashera CAT is denoted as FF(i), while FF(b) represents the total population of Ashera CAT. FFmax indicates the highest value of the FF, and FFmin refers to the lowest value of the FF.

Tracing Modes. In this phase, the trait of Ashera Cats is replicated by mimicking their tracing actions. Initially, random velocity values are assigned to each dimension of an Ashera Cat’s position. However, in subsequent iterations, the velocity values must be modified accordingly. This method of movement for Ashera Cats is described as pursues:

(i) Upgrade velocities (V (ASHERA CAT)) for all dimensions regarding to below Eq. (19):

$\begin{gathered}V(CAT)=V(CAT)+a * c\left(x\left(new_{-}cat\right)-\right. x\left(old_{-}cat\right))\end{gathered}$         (19)

where, a and c are constants.

4.5.2 Hyper parameter tuning process

The proposed Optimized are utilized to optimize the weights of X_LSTM dense networks. Initially, the hyperparameters are opted at random and moved to the X-LSTM training network. The novel FF which is coined based on ACO model is given in Eq. (20).

$\begin{gathered}\text {Fitness Function}=\operatorname{Min}(\operatorname{MSE}(\text {Predicted value -} \\ \text {Actual Value}))\end{gathered}$          (20)

The fitness function is computed based on the minimum error which is measured by MSE (mean Square Error) among the predicted value and actual value. Once the hyperparameters are optimized using Eq. (16), dense training layers classifies data into normal and attacks. The complete phase of operation of the recommended approach is represented in Algorithm-1.

4.6 FL model for the proposed network

FL is considered to be promising framework that used to construct the privacy-preserving learning models that guards the privacy. In the progress of learning framework, global model based on the ACAT-X-LSTM model is trained with the help of other participants and the decentralized data overseen by the central cloud/server. The individual receives a common global scheme from the server and execute training on their individual local datasets. Afterward, they pass the weights or gradients of their locally trained scheme to the task publisher for updating the global approach. (Algorithm-2 presents the working mechanism of the proposed model). Specifically, Fed-X-LSTM is formulated with the objective function is rewritten relied on the Fed-Avg functions which are represented as follows:

$f(w)=\sum_{j=1}^j \frac{N(i)}{N} * F(W)$                  (21)

The algorithm follows a straightforward approach, where $j$ portrays the total count of participants, and $n(i)$ portrays the count of training samples for the $j$-th participant. Initially, specific nodes are chosen within each batch for training across epochs. Subsequently, each node transmits its weight upgrades to the central server.

$w<----w<---\eta \alpha L(w, h)$                (22)

The server then gathers all the wt+1k values to compute the weighted average of the updated global wt+1, which is subsequently transmitted to each participant.

$(w)=\sum_{j=1}^j \frac{N(i)}{N} * W(j+1)$               (23)

4.7 HCL schemes for techniques

As discussed PPFL requires more security in sharing of the private data to train the federated model. Therefore, the chaotic preserving procedures are used in FL. In this research, chaotic procedures are utilized along with ABE for maintaining the privacy and encrypting the data from the nodes to server. To establish the chaotic principles, a heterogeneous combination of the Henon and HCL maps is employed in the proposed framework. The multi-HCL attractors are favoured over varied residing chaotic maps like circle, sine, logistic maps and tent due to their superior randomness and the capability to manipulate chaotic trajectories by adjusting initial phases.

4.7.1 HCL attractors

Dynamic systems that exhibit multi-HCL attractors often demonstrate more intricate behaviour compared to typical chaotic systems with single-HCL attractors. The equation governing the state space for an automatic chaotic system is expressed as:

$\dot{x}_1=-a x_1+b x_2 x_3$               (24)

$\dot{x}_2=-c x_2^3+d x_1 x_3$               (25)

$\dot{x}_3=e x_3-f x_1 x_2$               (26)

The above Eqs. (24)-(26) could be revised by the adding the hyperbolic equation $p_1 \tanh \left(x_2+g\right)$ which is given in below equations:

$\dot{x}_1=-a x_1+b x_2 x_3$               (27)

$\dot{x}_2=-c x_2^3+d x_1 x_3$               (28)

$\dot{x}_3=e x_3-f x_1 x_2+p_1 \tanh \left(x_2+g\right)$               (29)

Chaotic attractor is acquired when $a=2, b=6, c=6$, $d=3, e=3, f=1, p_1=1, g=2$ and the chosen initial factors are $\left[x_1(0), x_2(0), x_3(0)\right]=[0.1,0.1,0.6]$.

When the hyperbolic function is applied originally with a parameter value of g=−3 and the starting conditions [0.1,−0.1,−0.6], a double-HCL attractor is observed, as portrayed in Figure 4.

4.7.2 Henon maps-its principles of working

Henon Maps [33] are the disruptive quadratic and non-linear maps given by its characteristic equation.

$X_{n+1}=1-a X_n^2+Y_n$               (30)

$Y_{n+1}=1-b X_n$               (31)

The classical maps rely on the dual parameters an and b which has the values of a=1.4 and b=1.3. For the classical values, Henon map is chaotic. For the varied values of an and b, henon maps may exhibit the chaotic behavior which can be identified with the several times of iteration. Figure 5 represents the chaotic behavior of the henon maps using classical values.

In the proposed HCL techniques, HCL and henon maps are combined to form the ensemble techniques. The random output from the henon maps will be the input to the HCL attractors. The integration of the two maps leads to the high randomness outputs that can be used to create the strong keys against the multiple keys. Figure 6 shows the bifurcation diagram for the proposed HCL technique. The mathematically HCL is expressed by modifying the Eq. (30) and Eq. (31).

4.png

Figure 4. Non –linear behaviour diagram of multi-HCL attractors

5.png

Figure 5. Non –linear behaviour diagram of Henon MAPS

6.png

Figure 6. Non–linear behaviour diagram of HCL maps at the initial condition when a=0.3, b=0.1

$X_{n+1}=1-a X_n^2+Y_n+\mathrm{F}(\mathrm{s})$             (32)

$Y_{n+1}=1-b X_n+\mathrm{F}(\mathrm{s})$             (33)

where, F(s) is HCL maps.

4.7.3 ABE-HCL encryption technique

In this study, Cipher text policy based attribute relied encryption schemes (CP-ABE) based on the HCL techniques. For the encryption process, access policies (AP) are initiated by utilizing user attributes and the data can only be decrypted by the receiver if their attributes fulfil the AP's conditions. As the first step, public key(PK) and master key(MK) are constructed. By utilizing the AP and PK, a ciphertext is created. The secret keys are generated by utilizing mater keys and an attribute sets. To make the model as more resistant against the collusion attacks, the proposed study incorporates the heterogeneous chaotic encryption schemes to be more resistant against the collusion attacks. In the existing technique, attribute-based encryption (CP-ABE) procedures often rely on bilinear pairings of elliptic curves, known as pairing-friendly curves. Let G₁ and G₂ be cyclic groups of prime order p, and let e: G₁ × G₁ → G₂ be a bilinear map. In this proposed research, all the maps are generated based on the HCL technique which produces the more random keys which are unpredictable for tampering.

Encryption using HCL maps introduces additional security and privacy levels by modifying the input parameters (Algorithm-3). For encryption with HCL maps, a permutation operation is performed between each element of the input data and a chaotic value constructed by the HCL maps. The ith element of the plaintext data is diffused with the random value from the HCL maps to generate robust encrypted data. Prior to encryption, both the HCL maps and the data are scaled to a common factor of 16 to reduce process challenge. Similarly, in the reversible operation (Algorithm-4), a diffusion operation is performed among the encrypted data and the same encryption key (or parameter), which restores the original plaintext. The unique properties of chaotic systems, like ergodicity, sensitivity and determinism to initial phases, make them an attractive option for designing cryptographic systems. A prominent strength of chaos-relied encryption approaches is their computational efficiency [34].

The encryption and decryption processes, as detailed in Algorithm-3 and Algorithm-4, comprise several stages: 1) Key Generation: The keys are created by looping among varied originating criteria of the HCL maps. 2) Diffusion: This phase facilitates the interaction among the data and the HCL keys to generate the encrypted data. The primary goal of this research is to establish a multi-variable relationship among the original and the encrypted data. Furthermore, during encryption, numerous loops are applied to refresh the HCL maps and keys. Each iteration may modify the key to initiate additional randomness, thereby enhancing the security of the encryption process as per the AP. 4) The final encrypted outcome unveils increased Irregularity and greater Quantitative independence from the original data.5) The reversible progress of decryption is involved by checking the cipher text in accordance to AP if matches, decryption starts otherwise it ends.