IEEE Computational Intelligence Magazine - November 2021 - 71

data during training by prespecified predictions.
By employing crossed techniques,
data hiding can also be achieved.
The authors in [12] used a backdooring
technique to train a model that deliberately
outputs specific data with a trigger.
The specific data have no noticeable
impact on the primary task for which
the model is designed. In [13], a zero-bit
watermarking algorithm was designed
for neural networks using a frontier
stitching algorithm based on adversarial
perturbations. Wu et al. in [14] proposed
a digital watermarking framework for
deep neural networks in which the
trained neural network can accomplish
the original task while embedding a
watermark into the outputted images.
Recently, the authors in [15] proposed a
digital watermarking technique to prove
the ownership of the deep models using
selective weights for watermarking.
However, the abovementioned methods
are not designed for multiple receivers.
In some situations [16], e.g., command
delivery, a commander (the sender) plans
to transmit different instructions (secret
data) to multiple officers (the receivers).
For the consideration of security and efficiency,
the media carrying secret data
should be sent only once. In this case,
multiple instructions should be embedded
into the same media and sent to officers
(multiple channels). Therefore, a data
hiding scheme for multiple receivers is
Except for the intelligent tasks, these network models can
be used for data hiding which aims to transmit additional
data via digital media without causing serious distortion
on the cover media.
desirable. Moreover, these data hiding
methods for deep learning are based on
CNNs. Recently, a new network architecture
called CapsNets (capsule networks)
was developed [8]. The rudiments
of CapsNets were proposed in 2011 [17],
and enriched in 2017 with the routingby-agreement
mechanism [8]. In CapsNets,
several neurons are constituted as a
capsule that represents some properties
of an entity, e.g., pose, viewpoint, and
velocity. The capsules in a low-level layer
make pose predictions for the high-level
capsules using a transformation matrix.
Most importantly, a routing-by-agreement
mechanism is designed to pass the
information of the relevant capsules to
the next layer by iterative clustering.
Using this structure, variations of an entity
can be maintained. More details on
CapsNets will be given in Section II-B.
Currently, a CapsNet performs satisfactorily
in some popular image sets such as
the MNIST dataset [18], and is promising
for some intelligent tasks, such as multiobjects
detection [19]. In addition to
these conventional intelligent tasks, CapsNets
can be used for data hiding with
multiple receivers.
In this paper, the proposed scheme
achieves data hiding for multiple receivers
using CapsNets as an example, as
shown in Fig. 2. Additional data for
multiple receivers are embedded into
CapsNets. With CapsNets containing
additional data, a receiver can extract the
data for himself using his embedding
key. For other parts of additional data,
the receiver cannot affirm their existence
and therefore cannot extract them.
The security of data hiding can be guaranteed
this way since the existence of
additional data will not be exposed.
To the best of the authors' knowledge,
this is the first work focusing on data hiding
for multiple receivers using CapsNets.
In the proposed scheme, additional data
are embedded into CapsNets during the
training process instead of modifying the
network parameters after training. As a
result, the detection accuracy of CapsNets
decreases slightly. In addition, the
parameters of data decoding networks are
determined by embedding keys instead
Receiver 1
Sender
Embedding Key 1
Additional Data 1
CapsNets
Additional Data n
Data Extraction
Data Embedding
Embedding Key n
Embedding Key 1
Embedding Key n
CapsNets
Containing
Additional Data
Data Extraction
FIGURE 2 Data hiding for multiple receivers via CapsNets.
Additional
Data n
Additional
Data 1
Receiver n
NOVEMBER 2021 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE 71
IEEE Computational Intelligence Magazine - November 2021

Table of Contents for the Digital Edition of IEEE Computational Intelligence Magazine - November 2021
