First off, note that everything that a transmitter does for emissions, so too susceptibility is done to the receiver. The signal path couples symmetrically with the EMI environment, whether conducted or radiated. We say it obeys reciprocity. Now, the aggressor and victim generally aren't the same -- we generally aren't worried about a transmitter subject to incident EMI (unless it is very high level, like ESD), or the emission from a receiver (but that is its own interesting situation, when it sometimes happens*) -- but the coupling from external environment to the signal path is symmetrical, and so we expect similar degrees of attenuation from transmitter to far field, as from far field to receiver.
*For example, there might be an internal charge pump or something in a receiver (RS-232 interface ICs might be a typical example), or a radio set with a local oscillator that "bleeds out" of the RF front-end because the mixer isn't perfectly balanced (or isn't a balanced type at all, as was often the case for early UHF and microwave receivers). Back in the old days, TV viewership might've been measured, at least in part, by surveying the LO tones re-radiated by active sets via drive-by nondescript-white-van-studded-with-antennas methods. (I don't actually know how serious, or useful, this really was? I've heard of it, but not seen any articles actually discussing it; for the record, I consider it apocryphal but plausible.)
EMI is all about controlling what voltages manifest, what currents flow, where, and routing them away from signals, both so that our signals remain valid under specified adverse conditions, and so we aren't generating interference ourselves. On a mode fundamental level, it doesn't much matter if we're using a coax cable to carry a normal-mode signal, safely wrapped in ground, or a twisted-pair carrying a carefully balanced differential-mode signal. These are largely implementation differences -- and, most likely, cost is a decider between them.
We certainly must use like with like: sending a normal-mode signal (paired with simply GND) through a twisted pair, for example, puts an average V(signal)/2 on the cable, and thus considerable common-mode radiation. Likewise, an induced voltage in the ground loop between endpoints, half of that voltage is dropped across the (closely coupled) signal wire, thus we have very little noise margin (in the general case, anyway).
Comparing something like LVCMOS-level UART data sent over a twisted pair, or a plain unshielded multi-conductor cable, to a differential signaling standard like RS-422 over the same media, yes, we have a massive improvement: whereas the LVCMOS signal is delivering full-speed edges (in the, perhaps naive, unfiltered case, anyway), the RS-422 transmitter is delivering only the imbalance between its two pin outputs (typically specified to a level of 10s of mV), in the common mode. Even if we have imbalance along the cable (say because only one wire of the data pair happens to be on the outside of the multi-conductor cable), we might still have less than full peak-to-peak amplitude in the common mode, say when the cable is shorter than the edge rate of the transmitter (a short cable, relative to the edge rate, or to the maximum harmonics of the signal, means less opportunity for DM/CM mode conversion).
But like I said -- that's arguably a contrived example, because it's not placing like with like. If we sent that UART signal over a coax cable instead, or within a shielded multi-conductor cable, at worst we'd have coupling from that signal to other signals within the cable, but not to the outside world (well, not as much, that is -- shields aren't 100% effective, but 99.9+% effectiveness is achievable without too much work, and that covers a lot of situations already). Conversely (and within the same degree), induced currents flow exclusively on the outside of the shield, away from signals; as long as those currents are shunted around the circuit, from shield to shield, they remain separated from the signals and immunity remains excellent. But, "arguable" in that, if you simply didn't know better, you might assume a wire is a wire; and then you would discover first-hand what EMC can do to a signal. :)
Ethernet is another prime example of a twisted-pair medium. Back in the day, 10BASE2 was available: "thinnet", networking via 50Ω coax cable with multiple taps to serve a local network. ("Thin" in contrast to earlier networking schemes that used "fat" high impedance coax for, well, some reason; maybe just because driving signal currents into it was easier? Early networking standards were kinda wild!) This later competed with 10BASE-T, more or less the same thing using twisted pair, and then various improvements including to gigabit speeds and beyond. Today, twisted pair is king. The signaling is still very comparable to RS-422 or RS-485 differential (i.e., 100Ω impedance, 2.5V amplitude), but it's distinguished by somewhat lower signal levels (1000BASE-T, 1V), and more levels per symbol (100BASE-T uses three; 1000BASE-T, five; etc.). In any case, the transformers are what give Ethernet its robustness: not just extended common-mode range, but indeed isolation to 1.5kV; besides the transformers (which are balanced alright to begin with), there are common-mode chokes to attenuate the imbalance that the transformers inevitably have at high frequencies (near the transformer's cutoff frequency). Typical claims are 35dB CMRR -- not actually all that amazing, but certainly a vast improvement over an unbalanced, unshielded wire beaming that ~1V signal level into space.
Regarding single-ended media, yes, it definitely can be read differentially, and perhaps even often, should. Consider this example, an audio input circuit for an amplifier:
![enter image description here](https://cdn.statically.io/img/i.sstatic.net/FyLnj1nV.png)
The diff pair amplifier could be anything of course, I'd just drawn it here with a dual triode for vintage flavor. (On the upside, it does offer a pretty generous input CM range without any extra work -- check that 50V+ TVS!) The input would be an RCA jack, with shield/ground conspicuously exposed, so it must be used as not just the RF ground for EMC purposes*, but reference for the signal.
We could simply tie the shield to circuit ground, but we have a problem: if circuit ground is also earthed (for safety), we have one side of a ground loop wired up. If any connected equipment is also earthed, the voltage between earth connections will drop across these grounds. Copper braid isn't an effective shield below 10s of kHz, so this causes a voltage drop on the shield itself, for both modes (CM and DM), and we get ground loop noise -- usually hum or buzz. Lifting ground is a common act in audio systems -- but it doesn't have to be so dangerous; a better way exists.
RFI is a concern for amplifiers, particularly where (most often) the input stage acts to rectify incident RF, demodulating it into clicks or buzz that's audible on the output. For a vacuum tube amplifier, having several volts of input range, and little bandwidth, it's not much of a problem -- but it is just a matter of scale, and tubes will rectify some volts of RF; just as well, MOSFETs might rectify fractional to single-digit volts, and BJTs, 10s to 100s of mV. This diagram is of course a "belt and suspenders" sort of example, to show how much filtering you could employ, and where; but not necessarily that you need it for any given application. (It's rare that you hear RFI in a tube amp at all, and indeed you'd have to go to industrial levels of EMI, or worse, to see much effect. It is however very common to hear FM BCB, TV, or especially cell phone traffic these days, in bipolar circuitry; RFI filtering is not just recommended but actively necessary with them!)
For a somewhat contrary example, consider USB:
USB (2.0/compatible) is a data pair, and power and ground wire, in a shielded cable. The data pair carries either 3.3V LVCMOS (Low/Full Speed mode), or 400mV terminated to GND (High Speed), transmitted (mostly..!) differentially. The receiver -- at least as far as I know---I'd love to see real diagrams of receivers, but so far I've found nothing, if anyone has insight on this, comments are encouraged! -- is a normal CMOS input stage, combined with a low-voltage, high speed comparator. As such, it has clamp diodes from GND and to VDD, limiting the working common mode range to about a volt (give or take exact input thresholds). Or considerably less in HS, of course.
USB's receiver is open-circuit in Low/Full Speed mode, relying on source termination at the transmitter to achieve good signal quality. There is very little we can do with the pair at the receiver, considering 1. any filtering we do here, must load the differential signal (if the receiver is differential (as such) at all in this mode!), and 2. the receiver alternately transmits as well (both nodes on this point-to-point link are bidirectional) and we must not do anything that would interfere with that either. By extension, anything we do on one end, has to be done on the other by symmetry (or, if not done by any particular equipment we're connected to, we might at least have a reasonable expectation that they could).
Still, that's not awful, and practical examples exist -- RS-485 and CAN busses for example can be filtered moderately at each node on the bus. The filtering does impair the overall bandwidth or maximum fanout of the bus, but full occupancy is not often required, nor high bandwidth.
But USB has one sneaky catch, that often goes missed. It is not a differential standard at all, in fact. There is a single symbol in its line coding scheme which violates this, and intentionally transmits in the common mode: the SE0 ("single ended zero") symbol, used to delimit packets. If this symbol is filtered out, the USB link will fail to handshake, and throw an error.
Because of this, it is almost always better to firmly ground the connector housing/shield of the USB connector, to circuit ground. Now, for RFI purposes, this only needs to be an AC ground -- we can use multiple bypass capacitors in parallel for example (at some expense to maximum ESD/EFT immunity), to tie shield to PCB ground plane without making a galvanic (DC / low frequency) connection. There are cases where it's acceptable, or preferable even, to tie the shield to a metallic enclosure (when present), but most of the time, the PCB is the enclosure as such (rather, the nearest approximation to one), and also, I'm not aware of any USB hosts that don't actually tie GND wire to shield anyway, so it's doubtful the galvanic isolation would be useful in practice.
That is, to make the point clear: to read USB differentially, as in reading the data pair with respect to its reference plane (the cable shield), we would have to float the entire transceiver to that (shield) voltage, and then we'd have to find a solution to bring the transceiver's opposite side down to circuit ground (PCB reference plane). We might not be able to use the USB5V either. (Not that that's insurmountable -- USB data isolators are readily available, as are DC-DC converters. And if our interface is something simple like a USB-UART chip, UART is pretty easy to isolate. But these steps all add cost!)
And, to put a finer point on the concept: it is almost always wrong and objectively worse to "lift" a shield/ground from one end of a cable -- in unconditional terms, i.e. including at AC. Simply put, signal currents flowing on the outside of the shield, have no choice but to wrap around and go inside the cable at the cut. And then it's like you're using no shield at all, sure you get the shielding from inhomogeneous local fields (so they don't induce into some twists of the pair more than others), but you broke the ground loop that is specifically shunting RF noise currents away from your precious signals. It is regrettably common to see "ground coax at only one end" quoted in literature, forums posts, even here, but it goes to show how poorly understood in general EMC topics are. (And, to be fair, this stuff is hard. It's very complex, everything affects everything else. It's almost impossible to find any hard and fast rules, simple and invariant. It always depends on the environment and application, and something that worked well in one place might fail in another.)