Content uploaded by Stephan Pachnicke
Author content
All content in this area was uploaded by Stephan Pachnicke on Feb 18, 2022
Content may be subject to copyright.
Secure and Robust Federated Learning for Predictive Maintenance in
Optical Networks
Khouloud Abdelli1,2, Joo Yeon Cho1, and Stephan Pachnicke2
1ADVA Optical Networking SE, Fraunhoferstr. 9a, 82152 Munich/Martinsried, Germany
2 Christian-Albrechts-Universität zu Kiel, Kaiserstr. 2, 24143 Kiel, Germany
{KAbdelli, JCho}@adva.com, stephan.pachnicke@tf.uni-kiel.de
Abstract
Machine learning (ML) has recently emerged as a powerful
tool to enhance the proactive optical network maintenance
and thereby improves network reliability and reduces un-
planned downtime and maintenance costs. However, it is
challenging to develop an accurate and reliable ML model for
solving predictive maintenance tasks (e.g., anomaly detec-
tion, fault diagnosis, remaining useful prediction etc) mainly
due to the unavailability of a sufficient amount of training
data since the device failure does not occur often in optical
networks. Federated learning (FL) is a promising candidate
to tackle the aforementioned challenge by enabling the devel-
opment of a global ML model using datasets owned by many
vendors without revealing their business-confidential data.
While FL greatly enhances the data privacy, it is vulnerable
to various model inversion and poisoning attacks. In this pa-
per, we propose a robust collaborative learning framework
for predictive maintenance in a cross-vendor setting, whereby
the defensive mechanisms to protect against the aforemen-
tioned attacks are implemented. The multi-party computation
(MPC)-based secure aggregation is adopted to defend against
the model inversion attacks whereas a trained autoencoder
based anomaly detection model is used to recognize the
model poisoning attacks launched by compromised vendors.
The proposed framework is applied to the semiconductor la-
ser degradation prediction use case. We conduct experiments
on semiconductor laser reliability data obtained from differ-
ent laser manufacturers under various attack scenarios to
evaluate the attack defense and detection capabilities of the
proposed approach. Our experiments confirm that a global
ML model can be accurately built with sensitive datasets in
federated learning even when a subset of vendors is compro-
mised.
Introduction
Optical fiber networks compose the core of the telecommu-
nication infrastructure today due to their high capacity of
data transmission. Optical networks rely on fully functional
hardware components that run under optimal conditions. In
Copyright © 2022, Association for the Advancement of Artificial Intelli-
gence (www.aaai.org). All rights reserved.
order to reduce the risk of unplanned network interruption
and service outage, it is important to predict the degradation
of hardware network components correctly using analyzing
tools and techniques, by which the maintenance budget and
resources are allocated efficiently and timely. Due to the
great benefits in industry, the global predictive maintenance
market is expected to reach more than 13 billion US dollars
by 2026 (ReportLinker. (2021)).
Machine learning (ML) based techniques have emerged as
efficient tools to improve the accuracy of predictive mainte-
nance in the manufacturing industry and communication
networks. An ML model is trained by the historical data of
hardware failure and then the upcoming maintenance is pre-
dicted by real-time data gathered through measurement at
the edge. ML techniques can be useful, if a sufficiently
large, diverse, and realistic set of training data exists. Since
an ML model relies so heavily on good training data, the
availability of such datasets is a crucial requirement for this
approach.
However, it is challenging to develop a high-precision ML
model for predictive maintenance mainly due to the lack of
training data. The hardware failures or maintenance events
do not occur frequently so that it takes time until good and
meaningful training data are collected through the network.
Hence, an ML model is often trained using the accelerated
aging test results (e.g., a life cycle under the extreme tem-
perature or the over-powered condition) that are conducted
by hardware manufacturers. Since the components of net-
work equipment are usually produced by small and medium-
sized companies, such an ML model is trained based on the
limited amount of data that are owned by each manufacturer.
This situation can be relieved, if the training dataset can be
aggregated from multiple vendors and consolidated in a cen-
tral location. Since collaborative learning allows to train a
model on larger datasets rather than the dataset available in
a single vendor, a higher quality and more accurate ML
model can be built. However, such collaboration is not
straightforward in reality since vendors are not willing to
share their training data with external companies. Aging test
data are often company-confidential and trade secret. More-
over, sharing data with foreign companies may be prohib-
ited by privacy protection regulations in their home coun-
tries. To overcome such data-privacy concerns, federated
learning (FL) (i.e., collaborative learning) has been pro-
posed by enabling many vendors to collaboratively train a
global ML model without sharing their local private data
with others. However, FL is susceptible to various attacks
such as inversion model attacks aiming to compromise the
data’s confidentiality and poisoning attacks preventing the
global model from converging and thereby adversely im-
pacts its performance.
In this paper, we propose a secure and robust collaborative
learning framework incorporating defensive mechanisms to
defend against above attacks, using cross-vendor datasets
for predictive maintenance in optical networks. We apply
our approach to the use case of predicting the degradation of
semiconductor laser devices deployed in optical networks.
The experiments are performed using laser reliability data
from different laser manufacturers under various attack sce-
narios to test the efficiency of our defensive mechanisms in
protecting against attacks launched by compromised ven-
dors.
The rest of this paper is structured as follows: Section 2
gives some background information and related work. Sec-
tion 3 presents the proposed framework as well as the de-
fending mechanisms involved in the framework. Section 4
describes the validation of the presented framework using
experimental data. Conclusions are drawn in Section 5.
Background and related work
Federated Learning
Federated Learning (FL) is a framework of enabling distrib-
uted parties to work together to train machine learning mod-
els without sharing the underlying data or trusting any of the
individual participants (Bonawitz, et al., 2017). FL can be
used to build an ML model from various companies for the
purpose of predicting the failures, repairs, or maintenance of
network systems. With the FL technique, the training data is
not required to be centralized, but can instead remain with
the data owners. Each vendor trains an ML model on their
private data and using their own hardware. These models are
then aggregated by a central server (e.g., a network operator)
to build a unified global model that has learned from the pri-
vate data of every vendor without ever directly accessing it.
Hence, confidential training data (e.g., aging test results of
products) are not visible to a server, nor other competitive
vendors. An important challenge in FL is to prevent a server
or other vendors from reconstructing the private data of any
vendor while collaborating at any circumstances. While a
secure aggregation protocol in FL addresses the strong pri-
vacy of the data of the vendors, the FL framework creates a
new attack surface during the model training process. Since
the vendors have full control over local training processes,
they may submit arbitrary updates to change the global
model without being detected. Among the broad range of
attacks on FL, the following attacks are the most relevant to
our use case:
Model inversion attack
An attacker can intercept the updated local models and ex-
tract the private training data from the models. For example,
in (Fredrikson, Jha, & Ristenpart, 2015), the authors demon-
strated a model inversion attack that could extract images
from a face recognition system, which look suspiciously
similar to images from the underlying training data.
Local model poisoning attack
This attack injects poisoned instances into the training data,
or directly manipulates model updates during the aggrega-
tion protocol. An attacker can compromise some vendors
and thereby he may upload the poisoned local models,
which are highly deviating from the global model. As a re-
sult, the attacker can tamper with the weights of the global
model or inject a backdoor into it, misclassifying specific
inputs into the target class as intended by the attacker.
Secure aggregation
Secure aggregation in FL is a cryptographic protocol that
enables each vendor to submit a local model securely and a
server learns nothing but the sum of the local models. A se-
cure aggregation method for mobile networks was presented
in (Bonawitz, et al., 2017) and (Bell, Bonawitz, Gascón,
Lepoint, & Raykova, 2020). This method relies on a pair-
wise secret exchange and Shamir's t-out-of-n secret sharing
scheme, focusing on the setting of mobile devices where
communication is extremely expensive, and dropouts are
common.
There is a rich literature exploring secure aggregation in
both the single-server setting (via additive masking
(Bonawitz K. A., et al., 2016), via threshold homomorphic
encryption (HE) (Halevi, Lindell, & Pinkas, 2011), and via
generic secure multi-party computation (MPC) (Burkhart,
Strasser, Many, & Dimitropoulos, 2010) as well as in the
multiple non-colluding servers setting (Corrigan-Gibbs &
Boneh, 2017). For instance, one can perform all computa-
tions using a fully homomorphic encryption scheme result-
ing in low communication but very high computation or us-
ing classical MPC techniques with more communication but
less computation. Other works use a hybrid of both and thus
enjoy further improvement in performance (Mishra,
Lehmkuhl, Srinivasan, Zheng, & Popa, 2020) (Juvekar,
Vaikuntanathan, & Chandrakasan, 2018). Nevertheless, it is
still an open question how to construct a secure and robust
aggregation protocol that addresses all the challenges.
Autoencoder
An autoencoder (AE) is a type of artificial neural network
seeking to learn a compressed representation of an input in
an unsupervised manner (Kramer, 1991). An AE is com-
posed of two sub-models namely the encoder and the de-
coder, whereby the former is used to compress an input
into lower-dimensional latent-space representation
through a non-linear transformation, and the latter maps
the encoded representation back into the estimated vector
of the original input vector as follows:
where and represent the activation functions of the en-
coder and the decoder respectively. The weight matrix
(resp. ) and bias vector (resp. ) are the learnable pa-
rameters for the encoder (resp. decoder).
The training objective of the autoencoder is to minimize the
reconstruction error between the output
and the input ,
referred as the loss function , typically the mean square
error (MSE), expressed as:
where denotes the set of the parameters
to be optimized.
AE has been widely used for anomaly detection by adopting
the reconstruction error as anomaly score. It is trained with
only normal data representing the normal behavior. After
training, AE will reconstruct the normal instances very well,
while it will fail to reproduce the anomalous observations by
yielding high reconstruction errors. The process of the clas-
sification of an instance as anomalous/normal is shown in
Alg. 1.
Gated Recurrent Unit
The Gated Recurrent Unit (GRU) recently proposed by
(Cho, et al., 2014) to solve the gradient vanishing problem,
is an improved version of standard recurrent neural net-
works (RNNs), used to process sequential data and to cap-
ture long-term dependencies. The typical structure of GRU
contains two gates namely reset and update gates, control-
ling the flow of the information. The update gate regulates
the information that flows into the memory, while the reset
gate controls the information flowing out the memory.
The GRU cell is updated at each time step t by applying the
following equations:
(4)
(5)
(6)
(7)
where denotes the update gate, represents the reset gate,
is the input vector, is the output vector, and repre-
sent the weight matrix and the bias vector respectively. is
the gate activation function and tanh represents the output
activation function. The “” operator represents the Hada-
mard product.
Related work
In (Bonawitz K. , et al., 2017), a practical secure aggregation
technique in an FL setting was proposed over large mobile
networks. Such a framework does not fit for our use case
due to multiple reasons. Firstly, in our use case, a global
model is not shared with data owners (vendors). Each ven-
dor gets a benefit by receiving an individual maintenance
result (e.g., the difference between the prediction and the
real failure) after the global model is deployed and hardware
degradation is predicted. Secondly, the scalability is not im-
portant since the number of vendors is not very large and
dropouts are expected to be rare. On the other hand, secure
aggregation is critical since the disclosure of the private
training dataset may give negative impact on the data own-
er's business.
Another interesting work on collaborative predictive
maintenance was presented in (Mohr, Becker, Möller, &
Richter, 2020), where a combination of blockchain and fed-
erated learning techniques was applied. We apply a multi-
party computation technique for data privacy since it is more
suitable for our use case. More recently, in (Zheng, et al.,
2021), an end-to-end platform for collaborative learning us-
ing MPC is proposed. Though it is an interesting approach,
it is unlikely that this platform can be applied to our use case
since the collaborative learning through the use of release
policies and auditing is not preferable to the predictive
maintenance.
Algorithm 1: Autoencoder based anomaly detection
Input: Normal dataset , anomalous dataset
threshold
Output: reconstruction error
1: train an autoencoder given the normal data
2: for to do
3:
4: if then
5: is anomalous
6: else
7: is normal
8: end if
9: end for
Proposed Framework
Figure 1 illustrates the proposed secure collaborative learn-
ing framework for predictive maintenance in optical net-
works. We consider a FL approach that assumes N vendors
for collaborative training of a global ML model under the
control of an aggregator server hosted by an optical network
operator, while keeping every client’s data private. Each
vendor builds a local model using its own training dataset
and uploads it to the server. The private dataset remains in
the vendor's domain and is never exposed to other compa-
nies. The local model updates are sent securely to the server.
At the server side, an anomaly detection method adopted to
defend against the local model poisoning attacks is used to
firstly recognize the abnormal local model updates sent by
potentially compromised vendors, which are discarded. Af-
terwards, a server builds a global ML model by aggregating
only normal local ML models iteratively and averaging
them to form an updated global model proportional to the
size of dataset. An MPC-based secure aggregation defend-
ing against the model inversion attack is adopted. In our
framework, a secure aggregation protocol is tolerant to the
malicious behavior of participants in an honest-majority
model; that is, a server and majority of vendors are assumed
to be honest, yet some may be malicious or unreliable. Using
the global model, the potential risk of hardware failure or
degradation and corresponding maintenance events are pre-
dicted, and the necessary resources are proactively prepared
to run optical networks without disruption. Compared to the
original FL, the local models are not many, and the dropouts
are very rare in our framework. Furthermore, an updated
global model is not shared with vendors. The reason is that,
while a global model is a valuable asset to the network man-
agement, it is not really beneficial to the vendors. Instead,
each vendor receives the personalized maintenance report
which contains the discrepancy between its local model and
the global model, which is useful to improve the quality of
products in the future.
MPC-based Secure aggregation
Suppose that the server and vendors (clients) behave hon-
estly, but curiously (semi-honest model). That is, all partic-
ipants follow the protocol exactly as instructed, but also try
to retrieve the private data of other vendors, if possible. Un-
der this scenario, a simple n-out-of-n additive secret sharing
scheme can be used to prevent the model inversion attack as
well as keep the privacy of local models.
Suppose is the number of clients, and each client has its
own local model
where . The client gener-
ates a random linear mask and sends
to the server.
Also, the client divides into additive shares
,…,} in such a way that
.
Note the size of is similar to those of shares. These
shares are distributed to other clients in such a way that each
client receives a unique share out of shares. In result, the
client receives
Finally, the client sends the sum of the shares
to the server. This process is repeated for all clients. By ag-
gregating one-time padded local models and the sum of the
shares, the server can calculate the sum of the local models
as follows:
where denotes the size of the data of the client . repre-
sents the size of the aggregated data of all the clients (
)
An overview of the secure collaborative learning procedure
is shown in Fig. 2.
Autoencoder based anomaly detection
An autoencoder based anomaly detection method is adopted
to detect and exclude malicious local models updates from
the aggregation process. It is used to compute the recon-
struction errors of the local model updates. If the reconstruc-
tion errors are high, the model updates are considered mali-
cious and thereby removed.
The autoencoder is trained with a dataset
incorporating the local model updates (i.e.,
Fig. 1. ML-based predictive maintenance process in a dishonest
setting.
Fig. 2. Secure collaborative learning using Secret Sharing in FL.
model weights) sent by trusted clients (i.e., normal weights)
under no attack setting and stored at the server. The dimen-
sionality of the model weight is reduced to produce a
low-dimensional input in order to reduce the computational
complexity due to the high dimension of the model weight.
The generated input is fed then to the autoencoder for train-
ing, whereby the encoder compresses the input into a lower-
dimensional latent vector which is then reconstructed by the
decoder.
After the training phase, the autoencoder is able to recognize
the normal weights and mark any weight that deviates from
the data seen during the training as an anomaly. The recon-
struction error between the input weight and the recon-
structed weight is used as an anomaly score. If the anomaly
score exceeds a pre-defined threshold, the weight is recog-
nized as anomalous potentially sent by a malicious client,
and thereby it is removed and not considered for the update
of the global model. The threshold is optimized in order to
improve the detection capability of the autoencoder for dif-
ferent poisoning model attacks.
Validation of the Proposed Framework
Use case: Optical Transmitter Degradation Prediction
Semiconductor lasers are considered as one of the most
commonly used optical transmitters for optical communica-
tion system thanks to their high efficiency, low cost, and
compactness. They have been rapidly evolved to meet the
requirements of the next generation optical network in terms
of speed, power consumption, etc. However, during opera-
tion, the performance of the laser can be adversely impacted
by several factors such as contamination, crystal defects,
facet oxidation etc. Such factors are hard to predict, and their
interaction can lead to complex degradation mechanisms
which are hard to model. The semiconductor laser degrada-
tion occurs in three different modes: rapid, catastrophic, and
gradual. Each degradation mode is characterized by its own
signature depending on the laser’s architecture and compo-
sition. Among the degradation models, a catastrophic mode
is considered as the most challenging and hazardous ons as
it appears as a quick and sudden failure after a normal oper-
ation of the device. Therefore, it is hard to predict such deg-
radation leading to the end of the life of the laser, and
thereby resulting in optical network disruption and high
maintenance costs. Therefore, it is highly beneficial to pre-
dict the degradation of the semiconductor laser device after
its deployment in optical communication system in order to
enhance the system reliability and minimize the downtime
costs.
ML techniques could provide a great potential to tackle the
laser degradation prediction problem (Abdelli, Griesser, &
Pachnicke, 2020). The development of such prognostic
methods requires the availability of run-to-failure data sets
modelling both the normal operation behavior and the deg-
radation process under different operating conditions. How-
ever, such data is often unavailable due the scarcity of the
failures during the system operation and the long time re-
quired to monitor the device up failing and then generating
the reliability data. That is why accelerated aging tests are
often adopted to collect run-to-failure data in shorter time by
speeding up the device degradation by applying accelerated
stress conditions resulting in the same degradation process
leading to failure.
However, the burn-in aging tests are carried out just for few
devices due to the high costs of performing such tests.
Hence, the amount of the run-to-failure data that can be de-
rived from such tests, might be small, which can adversely
affect the performance of ML model (Abdelli, Griesser, &
Pachnicke, A Hybrid CNN-LSTM Approach for Laser
Remaining Useful Life Prediction, 2021). Therefore, an FL
approach is considered as a promising candidate to address
the aforementioned problem, whereby different semicon-
ductor laser manufacturers (i.e vendors) collaborate with
their small local dataset, stored at their premise, in order to
build an accurate and reliable global laser degradation pre-
diction model with good generalization and robustness ca-
pabilities.
Note that the semiconductor laser manufacturers might have
different types of laser devices with various characteristics
leading to different degradation trends, and that the data
owned by each vendor is derived from aging tests conducted
under different operating conditions. State that the global
model is running on a server hosted by an optical network
operator owning the infrastructure in which the semiconduc-
tor lasers manufactured by the different vendors are de-
ployed.
We consider an FL system composed of a server and
N
cli-
ents (i.e., vendors) that collaboratively train a global model
to predict the semiconductor laser degradation using the Fe-
dAvg algorithm (McMahan, Moore, Ramage, Hampson, &
Arcas, 2017).
The clients securely send the local model weight updates to
the server using MPC. A GRU based model is used as global
model to solve the task of semiconductor laser degradation
prediction. A convolutional autoencoder implemented at the
server is adopted as an anomaly detection method to detect
the anomalous weights sent by the malicious clients.
Experimental data
To evaluate our FL framework, we adopt different datasets
obtained from semiconductor laser manufacturers. The da-
tasets represent the reliability data of two different types of
semiconductor lasers namely vertical-cavity surface-emit-
ting laser (VCSEL) and tunable distributed feedback (DFB)
laser. VCSEL and DFB lasers differ in semiconductor ma-
terials and resonator structures, and are characterized by dif-
ferent degradation trends. Each dataset is derived from var-
ious accelerated aging tests performed according to
Telcordia GR-468 CORE requirements for multiple devices
with various characteristics (e.g., VCSELs with different
oxide aperture sizes…) operating under several operating
conditions and carried out under high temperature (50°C
) to strongly increase the laser degradation and
thereby speed up the device failure. Depending on the oper-
ating conditions, the duration of the aging tests is varied
(i.e., 2000h, 3000h, 3500h, 15000h). The output power (i.e.,
degradation parameter) is monitored under constant operat-
ing current . The failure criterion of the device is defined as
the decrease of the output power by 1 dB (20%) of its initial
value. Figure 3 shows examples of aging tests results of
semiconductor lasers conducted under different operating
conditions. As depicted in Fig. 3, few VCSELs are degraded
or failed during the aging tests, whereas more tunable DFB
lasers exhibit degradation.
In total, a dataset of 6,564 samples incorporating 8-length
sequences composed of monitored output power measure-
ments of length 6 combined with the operating conditions
namely and , is built. We assign to each sample the state
of the device (normal or degraded). For training and testing
the ML model to early predict the laser degradation, we con-
sider the samples of early degraded devices (i.e., during the
first stage of degradation, exhibiting a decrease of output
power of value between 5% and 10%). The said data is then
normalized and randomly divided into a training data (com-
prising of 80% of the samples) and a test dataset (the re-
maining 20% for testing). The training data is split then into
= 5 clients with different parts. Note that each client owns
a data of either different types of lasers than the other clients
or same type of lasers but manufactured by different laser
manufacturers (i.e., different materials and structure) and
tested on different wafers, leading to heterogeneous feder-
ated setting.
Global model
The adopted ML model to predict the degradation of the
semiconductor laser is a GRU-based model as GRU is good
at processing sequential data and to capture the relevant fea-
tures underlying the laser degradation trend under different
operating conditions. The architecture of the GRU model is
composed of one GRU layer containing 25 cells. The GRU
model takes as an input the sequence of length 8 including
the output power measurement values collected till time
combined with and , and outputs the state of the device
(“normal” or “degraded”) at the prediction time . The train-
ing of the global model is carried out in an iterative process
as follows:
• The server distributes the global model to
clients.
• Each client k trains the model locally using its local
data Dk and updates the weight for epochs
of Adam with mini-batch size of to compute
.
• The server securely aggregates each client’s
using MPC.
• An autoencoder-based anomaly detection method
is used to detect anomalous weights sent by the cli-
ents.
• The update of the global model .is computed
by a weighted averaging of only normal weights.
The above-described process is repeated for multiple com-
munication rounds (e.g., number of aggregation) to
improve the performance of the global model. For our ex-
periments, , and are set to 8, 10 and 20, respec-
tively.
Anomalous weight detection Method
A convolutional autoencoder implemented at the server is
used to identify the anomalous weights and thereby detects
the potentially malicious clients. The model contains an en-
coder and a decoder sub-model with 5 layers. The encoder
takes an input
(d)
(c)
(b)
(a)
Fig. 3. Experimental aging test data of semiconductor lasers con-
ducted under different operating conditions: (a) aging tests of
VCSELs with different oxide aperture sizes performed at 85°C, (b)
aging tests of tunable DFB lasers conducted at 90°C, (c) aging tests
of VCSELs performed at 50°C, and (d) aging tests of VCSELs con-
ducted at 70° C.
takes as an input a vector of length 75. It encodes the input
into low dimensional features through a series of 2 convolu-
tional layers containing 64 and 32 filters of size 3 x 1 with a
stride of 1. The decoder is inversely symmetric to the en-
coder part. It consists of 3 transposed convolutional layers
used to up-sample the feature maps. The last transposed con-
volutional layer with 1 filter of size 3 x 1 selected as an ac-
tivation function for the hidden layers of the model. The loss
function is set to the MSE, which is adjusted by using the
Adam optimizer.
Experimental Results
Prediction Capability Evaluation
The performance of the proposed FL framework is com-
pared to two baseline models including a model trained by
applying a traditional centralized approach and a locally
trained model without participating in the FL approach (i.e.,
localized model). The centralized approach is trained with
the data from all the clients, which is collected and stored at
a single server. The localized model is trained on the client’s
premises without model sharing during the training proce-
dure. The different approaches are evaluated using as eval-
uation metrics the accuracy, the precision, quantifying the
relevance of the predictions made by the ML model, the re-
call (i.e sensitivity), providing the total relevant results cor-
rectly classified by the ML model, and the F1 score, the har-
monic mean of precision and recall. The results of the com-
parison shown in Fig. 4 demonstrate that first the FL frame-
work outperforms the localized model by providing 11.4%,
9.62%, 14.7% and 12.24% improvements in accuracy, pre-
cision, recall and F1 score metrics, respectively, and second
that the FL approach achieves similar performance as the
centralized approach while ensuring data privacy.
Figure 5 shows the states of some tested devices predicted
by the ML model trained using the FL approach by giving it
as input the output power measurements monitored till
5,000 h (i.e., time of prediction). As depicted in Fig. 5, the
ML model accurately and early predicts the degraded de-
vices before reaching the failure criterion, which proves the
usefulness of the adopted ML model in early predicting the
degraded/failed devices.
The length of the input sequence, specifically the length of
the sequence of the output power measurements, has a sig-
nificant impact on the degradation prediction capability of
the ML model. As shown in Fig. 6, increasing the length of
the sequence of the output power measurements helps the
ML model to capture more information about the degrada-
tion trend and thereby to achieve better degradation predic-
tion capability performance (i.e., yielding better accuracy,
precision, recall and F1 scores). However, rising the length
of the sequence too much (higher than 6) can lead to over-
fitting and thus reduces the performance of the ML model.
Robustness to model poisoning attacks
The anomalous weight detection model is compared to de-
fense-based methods namely krum (Blanchard, El Mhamdi,
Guerraoui, & Stainer, 2017), Trimmed Mean (Yin, Chen,
Kannan, & Bartlett, 2018) and Median. The testing accuracy
Fig. 4. Comparison of the federated (FL), centralized and local-
ized approaches.
Fig. 6. Impact of the output power sequence length on the perfor-
mance of the ML model.
Fig. 5. Assessment of early degradation prediction capability of
ML model.
achieved by the global model for each communication round
is adopted as evaluation metric. We consider the following
three adversarial attacks launched by 20% of clients (i.e.,
one compromised or malicious vendor) for each communi-
cation round :
• Additive noise attack: the compromised vendor
adds a Gaussian noise to the local model update
and set it as , where denotes the
original local model update or weight. is a vector
derived from a gaussian distribution of mean 0 and
standard deviation of 2 (i.e., standard deviation of
the normal weights).
• Sign flipping attack: the malicious vendor flips
the sign of the local model weight as = ,
where is a constant selected randomly from a
range from 1 to 5.
• Same value attack: the compromised vendor
sets its local model weight as =
, where is
a constant set to 2 and
denotes all-one vector.
The results shown in Fig. 7 demonstrate that the proposed
method significantly outperforms the defense-based ap-
proaches for the considered attack scenarios. It can be seen
that the proposed method converges faster under all the set-
tings and achieves similar performance as the FedAvg algo-
rithm without attack, which proves the effectiveness of the
anomaly detection model in detecting the anomalous
weights and in mitigating the impact of launched attacks. As
depicted in Fig. 7, the considered baselines are more robust
to the additive noise attack and not effective against same
value attack. The performances of the defense-based meth-
ods are worse as they are not effective in defending against
attacks for not identically and independently distributed
(iid) settings, and the fraction of the malicious clients which
is required by Krum and Trimmed Mean cannot be known a
priori in FL.
Let denote the testing accuracy achieved by the global
model trained under no attack setting for all the communi-
cation rounds, and present the testing accuracy obtained
by the global model trained under an attack launched each
communication round. The impact of an attack () is de-
fined as the reduction of the accuracy of the global model
due to the attack. It is expressed as .
Figure 8 illustrates that the proposed method is most robust
to the different attacks by achieving the smallest attack im-
pacts under all the considered attacks.
(b)
(a)
Fig. 7. Testing accuracy under different attack scenarios: (a) ad-
ditive noise attack, (b) sign-flipping attack, and (c) same value at-
tack.
Fig. 8. Attack impacts of different model poisoning attacks on FL
system with defense-based methods.
(a)
(b)
(c)
Conclusion
Optical networks require a high level of reliability and sus-
tainability. Machine learning techniques are expected to im-
prove maintaining such networks efficiently. We showed
that an accurate and reliable ML model could be developed
in collaborative learning without the disclosure of the cli-
ents' sensitive datasets even in a malicious setting. Our ex-
periments confirm that (i) the presented FL approach
achieves a good prediction capability similar to the one
yielded by the centralized approach, and (ii) the proposed
autoencoder based anomaly detection model is efficient in
recognizing the anomalous weights potentially sent by ma-
licious clients and outperforms the defense-based methods.
Acknowledgments
This work has been performed in the framework of the
CELTIC-NEXT project AI-NET-PROTECT (Project ID
C2019/3-4), and it is partly funded by the German Federal
Ministry of Education and Research (FKZ16KIS1279K).
References
Abdelli, K., Griesser, H., & Pachnicke, S. (2020). Machine
Learning Based Data Driven Diagnostic and Prognostic Ap-
proach for Laser Reliability Enhancement. 22nd Interna-
tional Conference on Transparent Optical Networks (IC-
TON).
Abdelli, K., Griesser, H., & Pachnicke, S. (2021). A Hybrid
CNN-LSTM Approach for Laser Remaining Useful Life
Prediction. Proc. of the Annual Conf. of the Prognostics and
Health Management Society (PHM).
Bell, J., Bonawitz, K. A., Gascón, A., Lepoint, T., & Ray-
kova, M. (2020). Secure Single-Server Aggregation with
(Poly)Logarithmic Overhead. Cryptology ePrint Archive,
Report 2020/704.
Blanchard, P., El Mhamdi, E. M., Guerraoui, R., & Stainer,
J. (2017). Machine Learning with Adversaries: Byzantine
Tolerant Gradient Descent. Advances in Neural Information
Processing Systems.
Bonawitz, K. A., Ivanov, V., Kreuter, B., Marcedone, A.,
McMahan, H. B., Patel, S., . . . Seth, K. (2016). Practical
Secure Aggregation for Federated Learning on User-Held
Data. http://arxiv.org/abs/1611.04482.
Bonawitz, K., Ivanov, V., Kreuter, B., Marcedone, A.,
McMahan, H. B., Patel, S., … Seth, K. (2017). Practical Se-
cure Aggregation for Privacy-Preserving Machine Learning.
(pp. 1175–1191). Proceedings of the 2017 ACM SIGSAC
Conference on Computer and Communications Security.
Burkhart, M., Strasser, M., Many, D., & Dimitropoulos, X.
(2010). SEPIA: Privacy-Preserving Aggregation of Multi-
Domain Network Events and Statistics. 19th USENIX Se-
curity Symposium Security.
Cho, K., Merrienboer, B. v., Gulcehre, C., Bahdanau, D.,
Bougares, F., Schwenk, H., & Bengio, Y. (2014). Learning
Phrase Representations using RNN Encoder-Decoder for
Statistical Machine Translation.
Corrigan-Gibbs, H., & Boneh, D. (2017). Prio: Private, Ro-
bust, and Scalable Computation of Aggregate Statistics.
Proceedings of the 14th USENIX Conference on Networked
Systems Design and Implementation.
Fredrikson, M., Jha, S., & Ristenpart, T. (2015). Model In-
version Attacks That Exploit Confidence Information and
Basic Countermeasures. Proceedings of the 22nd ACM
SIGSAC Conference on Computer and Communications
Security.
Halevi, S., Lindell, Y., & Pinkas, B. (2011). Secure compu-
tation on the web: Computing without simultaneous interac-
tion. (pp. 132--150). Advances in Cryptology -- CRYPTO
2011.
Juvekar, C., Vaikuntanathan, V., & Chandrakasan, A.
(2018). Gazelle: A Low Latency Framework for Secure
Neural Network Inference. http://arxiv.org/abs/1801.05507.
Kramer, M. A. (1991). Nonlinear principal component anal-
ysis using autoassociative neural networks. AIChE Journal.
Li, L., Xu, W., Chen, T., Giannakis, G. B., & Ling, Q.
(2018). RSA: Byzantine-Robust Stochastic Aggregation-
Methods for Distributed Learning from Heterogeneous Da-
tasets. http://arxiv.org/abs/1811.03761.
McMahan, H. B., Moore, E., Ramage, D., Hampson, S., &
Arcas, B. A. (2017). Communication-Efficient Learning of
Deep Networks from Decentralized Data.
Mishra, P., Lehmkuhl, R., Srinivasan, A., Zheng, W., &
Popa, R. A. (2020). Delphi: A Cryptographic Inference Ser-
vice for Neural Networks. 29th USENIX Security Sympo-
sium.
Mohr, M., Becker, C., Möller, R., & Richter, M. (2020). To-
wards Collaborative Predictive Maintenance Leveraging
Private Cross-Company Data. INFORMATIK 2020.
ReportLinker. (2021). Global Predictive Maintenance In-
dustry. https://www.reportlinker.com/p05799417/Global-
Predictive-Maintenance-Industry.html?utm_source=GNW.
Simon, J. (2021). Predictive maintenance with machine
learning on AWS.
Yin, D., Chen, Y., Kannan, R., & Bartlett, P. (2018). Byz-
antine-Robust Distributed Learning: Towards Optimal Sta-
tistical Rates. Proceedings of Machine Learning Research.
Zheng, W., Deng, R., Chen, W., Popa, R. A., Panda, A., &
Stoica, I. (2021). Cerebro: A Platform for Multi-Party Cryp-
tographic Collaborative Learning. 30th USENIX Security
Symposium.
